Research ArticleHow to Construct a Power Knowledge Graph withDispatching Data

Shixiong Fan1 Xingwei Liu1 Ying Chen 2 Zhifang Liao 2 Yiqi Zhao2 Huimin Luo2

and Haiwei Fan3

1Beijing Key Laboratory of Research and System Evaluation of Power Dispatching Automation TechnologyChina Electric Power Research Institute Haidian District Beijing 100192 China2School of Computer Science and Engineering Central South University Hunan 410000 China3State Grid Fujian Electric Power Co Ltd Fuzhou 350003 China

Correspondence should be addressed to Ying Chen 1074647728qqcom and Zhifang Liao zfliaocsueducn

Received 24 April 2020 Revised 7 May 2020 Accepted 26 May 2020 Published 14 July 2020

Academic Editor Chenxi Huang

Copyright copy 2020 Shixiong Fan et al 0is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Knowledge graph is a kind of semantic network for information retrieval How to construct a knowledge graph that can serve the powersystem based on the behavior data of dispatchers is a hot research topic in the area of electric power artificial intelligence In this paper wepropose a method to construct the dispatch knowledge graph for the power grid By leveraging on dispatch data from the power domainthismethod first extracts entities and then identifies dispatching behavior relationship patternsMore specifically themethod includes threesteps First we construct a corpus of power dispatching behaviors by semi-automated labeling And then we propose a model called theBiLSTM-CRFmodel to extract entities and identify the dispatching behavior relationship patterns Finally we construct a knowledge graphof power dispatching data 0e knowledge graph provides an underlying knowledge model for automated power dispatching and relatedservices and helps dispatchers perform better power dispatch knowledge retrieval and other operations during the dispatch process

1 Introduction

Smart grids have made important progress in the research andintegration of dispatch automation systems [1] According tothe relevant dispatch documents such as the power dispatchingcontrol rules and the experience of the dispatcher togetherwith the dispatching system data and the operation state of thepower grid the dispatcher judges whether the current oper-ation state of the power system needs dispatching and whatkind of dispatching behavior is to be executed In the actualpower dispatching scenarios power dispatching tasks are stillhighly dependent on the dispatcherrsquos business knowledge anddispatching experienceMost dispatchers only understand localbusiness knowledge [2] and cannot effectively respond to otherdispatching business or global business

With the continuous integration of multiple dispatchingservices the expertise of experts or dispatchers also needs to beintegrated tomeet the needs of simultaneously handling complexmultiservice power dispatching problems To provide dispatcher

reference for dispatchers digital power experts have written thetexts of power dispatch which summarize and contain all aspectsof the global power dispatch business Hence studying knowl-edge organization methods for global dispatching texts buildingknowledge models based on multiple dispatching behaviors andimplementing knowledge expressions that flexibly and clearlyexpress business logic will help to improve the degree of auto-mation of dispatching systems and provide global knowledgesupport for intelligent grid dispatching

0e text of power dispatching is characterized byknowledge-intensive and abundant knowledge types andit is a kind of unstructured data Compared with struc-tured data with strict format and specification the ex-pression mode of power dispatching text is more flexibleand is more difficult to read and understand So it isnecessary to explore a natural language processingmethod for the dispatching text and a behavioralknowledge organization method that is suitable for thecharacteristics of the power dispatching behavior

HindawiScientific ProgrammingVolume 2020 Article ID 8842463 10 pageshttpsdoiorg10115520208842463

To solve these problems we use the knowledge graph toorganize knowledge of power dispatching behavior Tradi-tional relational databases face the problems of repeateddata weak data relationships and difficult updates Com-paratively the knowledge graph organizes knowledge in agraph topology which is more in line with the structure ofthe power system It regards relationships as an importantknowledge element which can better describe knowledgeentities and their relationships such as power environmentdispatching roles and power dispatching behavior Itsknowledge storage and retrieval are more flexible First weanalyze the text of power dispatching behavior combinenatural language processing technology to build a powerdomain dictionary mine power domain phrases andidentify entities based on domain phrases and domaindictionaries to achieve coreference resolution 0en weanalyze the text characteristics of dispatching behaviordefine and organize the relationship of power dispatchingbehavior and use the graph structure to store entities andrelationships thereby constructing a knowledge graph ofpower dispatching behavior

0e rest of this paper is organized as follows Section 2introduces the related work Section 3 describes the con-structing method of knowledge graph of power dispatchingbehavior Section 4 introduces the experiment and evalua-tion results data set experimental design experimentaldetails and experimental results Section 5 introduces anexperimental summary and future work outlook

2 Related Work

Google first proposed the concept of the knowledge graph in2012 [3] which can formally describe things and their relatedrelationships in the real world 0e knowledge graph usesltentity relationship entitygt triples to store knowledge and ituses entities as nodes relationships as edges to build aknowledge network which conforms to the behavior rules ofgeneral subjects actions and action objects and uses graphstructures to describe the relationship At present many well-known knowledge graph projects organize a large amount ofdata extract knowledge from them for organization andmanagement and provide users with high-quality intelligentservices such as understanding the semantics of search andproviding more accurate search answers In recent years dueto the development of crowdsourcing [4] and open-sourceecosystems [5ndash7] the related research of constructingknowledge graphs by crowdsourcing and knowledge graphsin software has become a new research topic in the field ofknowledge graphs which also shows that knowledge graphsare flexible to organize domain knowledge

Entities are the basic units in the knowledge graph in-cluding attributes attribute values and the correspondencebetween related entities Wang [8] designed a named entityrecognition system based on text structure features 0ismethod needs to design the characteristics of text in differentfields separately which is not universal As the scale of datagrows the study on multilayer architecture and deep learningis extraordinarily important and necessary [9] In order toreduce manual rules and improve the generalization

capabilities of the model many methods based on deeplearning have been used for named entity recognition inrecent years For example Lample et al [10] proposed aneural network structure based on bidirectional LSTM andCRF for entity recognition 0is method does not rely onartificial features and domain-specific knowledge and hasexcellent versatility In the same period Chiu andNichols [11]proposed a bidirectional LSTM and CNN hybrid model forautomatic detection of the word and character-level featureseliminating the need for most feature engineering To over-come the problem of missing Chinese information whenusing the English relationship extraction method Han et al[12] generated a large-scale Chinese relationship extractiondata set based on the Chinese Encyclopedia and proposed anattention model based on the entity character features in theChinese relationship extraction method Leng and Jiang [13]proposed an improved SDAE model for entity-relationshipextraction 0is method eliminates the need to annotate therelationship manually and can extract the relationship be-tween the entities automatically through contextual features

0ere are many types of power grid equipment and therelationship between the types of equipment is complicated0ere is little research on knowledge graphs in the field ofpower grids Tang [14] merged existing multisource het-erogeneous power equipment related data to construct apower equipment knowledge graph to improve the datastorage efficiency and extraction process Given the problemof flattening and efficient utilization of power asset infor-mation Yang [15] proposed the general process of con-structing professional knowledge graphs in the power fieldand proposed a multisource heterogeneous power assetinformation fusion method based on knowledge fusion Liet al [2] based on the underlying data and business logicdata of the smart grid dispatching control D5000 systemused a top-down and bottom-up method to construct adispatch knowledge graph 0ese power knowledge graphconstruction methods based on power system data such asdispatch management systems face several challenges hugedata size complex data types diverse knowledge contentlow-value density and low data quality 0erefore the re-search on text data with higher value density has attractedmuch attention in recent years Wang et al [16] proposed amethod of entity recognition coreference resolution andrelationship extraction based on the records of defects inelectrical equipment and automatically constructed theknowledge graph of defects in electrical equipment to im-prove the retrieval quality of the records of defects inelectrical equipment Existing research usually processesunstructured text according to different text types andmodes in their application scenarios and most of them arenot universal Hence it is necessary to study and design aknowledge graph construction method for power dis-patching behavior based on power dispatching text

3 Methodology

In this section this paper will introduce the knowledgegraph constructing method of power dispatching behaviorand its technical roadmap is shown in Figure 1 First a

2 Scientific Programming

phrase extraction algorithm based on mutual informationand left and right entropy is used to extract power domainphrases construct a professional dictionary for power dis-patch and prepare a corpus of power dispatch behavior0en the BiLSTM-CRF model is built to train labeled dataand identify and extract entities in the power dispatchingdomain Finally by analyzing and summarizing the entityrelationships the power dispatching behavior relationshipsare extracted and a graph database is used to store andconstruct a knowledge graph structure

31 Corpus Construction Constructing a high-qualitydomain text corpus is a prerequisite for acquiringknowledge entities of power dispatching behaviorHowever there is a lack of labeled data in the field ofpower dispatching and manual labeling consumes muchenergy It is affected by the complexity of the domainentity category and the professionalism of the labelingpersonnel 0erefore this paper uses a phrase extractionalgorithm based on mutual information and left and rightentropy to get candidate phrases and selects and annotatesthem manually to get high-quality annotation data

311 Phrase Extraction Algorithm Based on Mutual Infor-mation and Left and Right Entropy Most of the griddispatching entities are nested combinations of multiplewords 0erefore in the traditional corpus labelingprocess the original corpus must first be segmented toclarify the boundary of the words which is convenient formanual labeling later Existing word segmentationtoolkits such as jieba word segmentation tools mostlyuse dictionary-based word segmentation methods 0edictionaries used are cross-domain general dictionariesmost of which commonly used vocabularies and lackprofessional vocabulary in the power field For examplethe word ldquooperation instruction ticketrdquo will be divided

into three words ldquooperationrdquo ldquoinstructionrdquo and ldquoticketrdquowhen using the general dictionary If we use the originalcorpus in the field of power dispatching to directlysegment words the effect of this method is not satis-factory 0erefore in the stage of the cold start of corpuslabeling in order to obtain labeled corpus this paper usesa novel unsupervised word discovery algorithm which isa phrase extraction algorithm based on mutual infor-mation and left and right entropy

0e algorithm first calculates the mutual informationbetween the words in the corpus 0e formula is as follows

PMIx y log2p(x y)

p(x)p(y) (1)

In Formula 1 p(x y) is the probability of two wordsappearing and p(x) is the probability of a single wordappearing We use specific examples to explain this 0ereare three dispatching behavior words ldquoProvincial Dis-patchingrdquo ldquoDispatcher on dutyrdquo and ldquoProvincial Dispatcheron dutyrdquo If the word frequency of ldquoProvincial Dispatchingrdquois 10 the word frequency of ldquoDispatcher on dutyrdquo is 20 andthe word frequency of ldquoProvincial Dispatcher on dutyrdquo is 5the total number of words is N and the total number ofdouble words is M then we have the following formula

PMI Provincial DispatchingDispatcher on duty

log25M

(10N)lowast(20N)

(2)

0e mutual information can reflect the relationshipbetween two words well 0e higher the mutual informationvalue is the higher the correlation between X and Y is themore likely X and y are to form phrases On the contrary thelower the mutual information value is the lower the cor-relation between X and Y is the more likely there is a phraseboundary between X and y

Knowledge graph ofpower dispatching behavior

BILSTM‐CRF model

Entity recognition

Entity and relationship data

Power dispatching texts

Natural language processingsuch as text segmentation

Domainphrase

Manual annotation

Phrase extraction algorithmbased on mutual informationand the left and right entropy

Labeledcorpus

Dispatchingbehavior entity

Relationshipdefinition

Relationshipextraction

Text processing

Relationship extraction

Figure 1 Technical roadmap for constructing a knowledge graph of power dispatching behavior

Scientific Programming 3

Mutual information indicates the relevance of the twowords Also we need to calculate the degree of freedom ofthe word 0e degree of freedom refers to the degree ofdiversity of adjacent words that appear on the left and rightsides of the word If the left and right sides of a candidateword are different words in different sentences the smallerthe connection between the word and other words thegreater the internal connection between the candidatewords that is the greater the possibility that the candidatewords have boundaries and are a single word

We use the left and right entropy to measure the degreeof freedom Entropy can describe information uncertaintyIn information theoretic learning correntropy has been awidely used nonlinear similarity measure method due to itsrobustness [17] 0e larger the left entropy and right entropyof a candidate the more uncertain the words that mayappear on the left and right sides of the candidate and thehigher the degree of freedom0e formula for calculating theleft and right entropy is as follows

EL(W) minus 1113944forallaisinA

P(aW | W) middot log2 P(aW | W)

ER(W) minus 1113944forallbisinB

P(Wb | W) middot log2 P(Wb | W)(3)

Taking the left entropy as an example suppose that theldquoDispatcher on dutyrdquo has several kinds of collocationsldquoNational Dispatcher on dutyrdquo ldquoProvince Dispatcher ondutyrdquo and ldquoTemporary Dispatcher on dutyrdquo then the leftentropy of the word ldquoDispatcher on dutyrdquo is as follows

minus EL(Dispatcher on duty)

P(N D |Dispatcher on duty)

middot log2 P(N D |Dispatcher on duty)

+ P(P D |Dispatcher on duty)

middot log2 P(P D |Duty Officer)

+ P(T D |Dispatcher on duty)

middot log2 P(T D |Dispatcher on duty)

(4)

0e final input result is the score of a series of words0ecalculation formula of the score is as follows

score PMI + min(left entropy right entropy) (5)

0ese scores are sorted from high to low We add the top100 words to the jieba word segmentation dictionary andthen perform word segmentation processing on the originalcorpus text to facilitate the manual labeling of the role ofwords in the later period

312 Manual Annotation Due to the fuzzy boundary ofChinese words and a large number of cross-nesting struc-tures in the grid dispatching entity the complexity of theidentification task increases Furthermore the data set in thispaper contains multiple categories of entities Consequentlyaccording to the word segmentation results obtained by theunsupervised phrase extraction method the word seg-mentation results need to be returned to the original corpus

after manual inspection 0en we use the BMESO labelingmechanism to convert it to the input format required by themodel and finally get a labeled Training data set 0e def-inition of the BMESO annotation model is shown in Table 1

0e labeling tool we use is YEDDA For named entitiesin the field of power dispatching we have summarized manydispatch documents and dispatch glossary classificationmethods 0en a number of collaborators form a team tocollaborate to review and determine the dispatching be-havior entities and finally they are classified as follows

(1) Scheduling mechanism (SM) including Chinarsquos fivemajor power generation groups regional powergeneration groups State Grid Corporation of ChinaRegional Power Grid Corporation managementorganizations and departments at all levels

(2) Scheduling personnel (SP) including leaders ofvarious organizations technical personnel at alllevels and dispatching personnel on duty at all levels

(3) Scheduling operation (SO) including but not limited todispatching operation related to the protection device

(4) Facilities (Fac) such as transformer bus line circuitbreaker switch knife gate protection device pri-mary equipment secondary equipment electricalequipment boiler equipment steam (water gas)turbine equipment power transmission equipmenttransmission equipment converter equipmentpower system chemical treatment and fueltransportation

(5) Management requirements (MR) including sched-uling management scope (equipment name)scheduling management mode and schedulinginstructions

(6) Electric power data (EPD) such as power-relateddocuments systems and operation tickets

(7) Scheduling condition (SC) the objective conditions forcertain dispatching under the power performance suchas the conditions for the power outage and powerstations or substations on both sides of the line

(8) Equipment state (ES) such as operation mainte-nance standby charging power transmissionpower failure and other equipment states

32 Entity Extraction In the past few years the rapid de-velopment of machine learning has attracted the attention ofmany researchers [18] In order to identify and extractknowledge entities this paper uses a Bidirectional LongShort-Term Memory (BiLSTM) model and ConditionalRandom Fields (CRF) model as a named entity recognitionmodel We use the annotated data of the annotated corpusabove for model training and extraction of knowledge en-tities in the field of power dispatching behavior0e BiLSTMmodel is composed of forward LSTM and backward LSTM0e LSTM model can memorize the long-term dependenceof sentences from front to back but it cannot encode in-formation from back to front Compared with a single LSTMmodel the BiLSTMmodel can obtain bidirectional semantic

4 Scientific Programming

dependence and obtain more comprehensive text infor-mation However the BiLSTM model does not guaranteethat the prediction results obtained at each output layer arecorrect and some prediction results that do not meet theconstraints of the training set may appear 0erefore theCRF model can be introduced to learn the constrainingrules thereby reducing the output of the model the prob-ability of an illegal sequence 0e annotated corpus con-structed above prepares for the building of an entityrecognition model At the same time annotated data is usedfor model training to identify entities with domainknowledge of power dispatching behavior

0e BiLSTM+CRF model is mainly composed of threelayers and the schematic diagram of the model is shown inFigure 2 0e first layer is the embedding layer 0e wordvector is trained by inputting the pretrained character vectorand word vector and the dictionary obtained in the previouscorpus labeling process is added to make the generated wordvector more capable of expressing semantics

0e second layer in the middle is the forward andbackward LSTM layer In order to make full use of wordmeaning and word order information the input sequence ofthe character vector and the word vector of the matchingdictionary are subjected to feature fusion through networkcalculation

0e BiLSTM layer automatically extracts sentence fea-tures uses the char embedding sequence (x1 x2 x3 xn)of each word in a sentence as the input of each time step ofBi-LSTM and then uses the hidden state sequence(h1rarr

h2rarr

h3rarr

hn

rarr) output by the forward LSTM and the

reverse LSTM (h1

larr h2

larr h3

larr hn

larr) 0e hidden state output

at each position is stitched by a position to obtain a completehidden state sequence (h1

rarr h2rarr

h3rarr

hn

rarr) isinRnlowastm

0e output of this layer is the score of each label of aword by selecting the highest label score as the label of theword

Finally the CRF layer is introduced for sentence-levelsequence annotation 0e parameter of the CRF layer is a(k + 2) times (k + 2) matrix A k is the number of labels in thelabel set and Aij represents the transfer score from the i-thlabel to the j-th label When labeling a location you can usethe label that has been labeled before0e reason for adding2 is to add a start state to the beginning of the sentence andan end state to the end of the sentence Adding the CRFlayer can consider the order between the labels of theoutput words of the Bi-LSTM layer adding some con-straints to the last predicted label to ensure that the pre-dicted label is legal

0is paper introduces the Dropout mechanism to pre-vent overfitting 0e Dropout mechanism prevents over-fitting by randomly deleting hidden neurons in the networkwith a certain probability0e neurons in the input layer andthe output layer of the network remain the same In this waythe hidden neurons deleted in each iteration cycle are dif-ferent which increases the randomness of the network andimproves the generalization ability of the network 0emodel code is shown in Algorithm 1

33 Relationship Extraction In order to mine the rela-tionship of power dispatching behavior this paper needs toanalyze the language characteristics of the relationship de-scription of power dispatching text Since the power dis-patching text is an unstructured natural language textwritten in Chinese it has the characteristics of the Chineselanguage grammar and the power field Its specific char-acteristics are as follows

0e sentence contains a large number of power domainentities In a sentence related to scheduling behavior theremay be three or more behavior subjects and objects at thesame time 0e relationship network formed by the rela-tionship between any two entities in the sentence is com-plicated However the entity-relationship category isrelatively straightforward and the entity relationships be-tween the restricted entity categories mostly belong to onecategory

Each sentence in the dispatching text corresponds to adispatch behavior and each segment corresponds to a typeof dispatch scenario with various types Understanding thedispatch statement requires professional knowledge ofelectricity and it is difficult for nonprofessionals to learn0e Chinese grammatical structure is more flexible andsophisticated than English with many grammatical phe-nomena such as condition sequence causality and passiveDifferent writers have different language habits and dif-ferent scheduling behaviors will also use different expres-sions At present there is a lack of available syntacticknowledge rule base in the field of electric power

0e dispatching text which is the basis of dispatchingbehavior is based on the summary of real-world dispatchingbehaviors 0e content is refined the data volume is

B-SP M-SP M-SP M-SP E-SP

P1

h1

P2 P3 P4 P5

h2 h3 h4 h5

h1 h2 h3 h4 h5

CRF layer

LSTM output

Backward LSTM

Forward LSTM

Embedding

Figure 2 Schematic diagram of the BiLSTM-CRF model

Table 1 0e definition of the BMESO annotation model

Annotation MeaningB 0e first word of the entityM 0e internal words of theE 0e suffix of the entityS Single entity wordsO Nonphysical constituent words

Scientific Programming 5

inadequate and there is a lack of labeled data 0e char-acteristics of multiple entities in the sentence make itchallenging to label entity-relationship data Machinelearning algorithms commonly used in the general fieldoften require large amounts of labeled data and cannot bedirectly applied to power dispatch texts

Based on the above characteristics we define the typesof power dispatching behavior relations as shown inTable 2 In the knowledge graph the edges representingthe relationship have directions and the relationshipedges in different directions may have different rela-tionship types

According to the above definition most of the two entitieshave only one type of relationship If two entities appear in ageneral sentence and their entity type meets the predefinedrelationship it can be considered that there is a predefinedrelationship between the two entities To extract the entityrelationship if there are multiple entities of the same type in asentence there may be a special relationship between theseentities such as a union When analyzing power dispatchingbehavior sentences words such as ldquocommonrdquo ldquoparallelrdquoldquoandrdquo and ldquoorrdquo are often used in the sentence to express theorder parallel and other relationships If there are relatedwords in the sentence that represent particular sentencepatterns such as juxtaposition negation and time it can bedetermined that the sentence has a special relationship and aparticular relationship type For the dispatcher and dispatchoperation entity the relationship type between the two typesof entities is judged according to the position characteristics ofthe entity in the sentence If the dispatcher entity is before thedispatch operation entity the relationship arrow is directed bythe dispatcher to the dispatch operation Otherwise the re-lationship arrow is determined by the dispatch0e operationis directed to the dispatcher 0erefore this paper sorts outand extracts the entity relations of power dispatchingbehavior

34 Knowledge Graph Construction and Retrieval Afterextracting power dispatching behavior entities and rela-tionships we use a graph database to store entity and at-tribute information and rely on entity relationships to

connect directed edges between entity nodes thereby con-structing a knowledge graph structure We use a graphdatabase query language to provide a retrieval method basedon knowledge graphs Neo4j database is one of the morepopular graph databases with good performance and afriendly user interface We use the Neo4j database as astorage database to construct a knowledge graph for powerdispatching and use the declarative graph query languageCypher provided by the Neo4j database for knowledge graphretrieval

4 Experiment

Based on the knowledge graph construction method pro-posed above this paper presents the experimental work oflabeling corpus construction knowledge entity extractionand knowledge graph construction of power dispatchingbehaviors with the power dispatch text data set In thissection we will detail the experimental design experimentaldetails and experimental results

41 Data Sets and Data Preprocessing In this paper wecrawled 29 documents related to power dispatching behaviorsuch as power grid dispatching procedures basic knowledgeof dispatching and disposal plans of dispatching failure0ese documents were written by professional power dis-patchers and these documents fully describe the powerdispatching business process dispatching requirements anddispatching behavior of dispatchers in the dispatchingprocess In this paper the above documents are used as theoriginal corpus for entity extraction and knowledge graphconstruction experiments In order to facilitate the follow-upwork we unify the document format remove the spaces andnumbers in the document and leave only character-typedata

42 Experiments and Result Analysis

421 Construction of Power Dispatching Behavior AnnotatedCorpus and Entity Extraction 0ere are a large number ofunlabeled entity vocabularies in the field of power grid

Input selfOutput Trained model(1) Initialize the model(2) Define the Embedding layer(3) Add the Embedding layer to the model(4) Add forward LSTM to the modelunits 128 return_sequencesTrue(5) Add Dropout(6) Add backward LSTM to the modelunits 64 return_sequencesTrue(7) Add Dropout(8) Add TimeDistributed layer to the model(9) Define the CRF layer and Add the CRF layer to the model(10) Parameter status of each layer of the output model(11) Return model

ALGORITHM 1 How to build BiLSTM+CRF named entity recognition model

6 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

phrase extraction algorithm based on mutual informationand left and right entropy is used to extract power domainphrases construct a professional dictionary for power dis-patch and prepare a corpus of power dispatch behavior0en the BiLSTM-CRF model is built to train labeled dataand identify and extract entities in the power dispatchingdomain Finally by analyzing and summarizing the entityrelationships the power dispatching behavior relationshipsare extracted and a graph database is used to store andconstruct a knowledge graph structure

31 Corpus Construction Constructing a high-qualitydomain text corpus is a prerequisite for acquiringknowledge entities of power dispatching behaviorHowever there is a lack of labeled data in the field ofpower dispatching and manual labeling consumes muchenergy It is affected by the complexity of the domainentity category and the professionalism of the labelingpersonnel 0erefore this paper uses a phrase extractionalgorithm based on mutual information and left and rightentropy to get candidate phrases and selects and annotatesthem manually to get high-quality annotation data

311 Phrase Extraction Algorithm Based on Mutual Infor-mation and Left and Right Entropy Most of the griddispatching entities are nested combinations of multiplewords 0erefore in the traditional corpus labelingprocess the original corpus must first be segmented toclarify the boundary of the words which is convenient formanual labeling later Existing word segmentationtoolkits such as jieba word segmentation tools mostlyuse dictionary-based word segmentation methods 0edictionaries used are cross-domain general dictionariesmost of which commonly used vocabularies and lackprofessional vocabulary in the power field For examplethe word ldquooperation instruction ticketrdquo will be divided

into three words ldquooperationrdquo ldquoinstructionrdquo and ldquoticketrdquowhen using the general dictionary If we use the originalcorpus in the field of power dispatching to directlysegment words the effect of this method is not satis-factory 0erefore in the stage of the cold start of corpuslabeling in order to obtain labeled corpus this paper usesa novel unsupervised word discovery algorithm which isa phrase extraction algorithm based on mutual infor-mation and left and right entropy

0e algorithm first calculates the mutual informationbetween the words in the corpus 0e formula is as follows

PMIx y log2p(x y)

p(x)p(y) (1)

In Formula 1 p(x y) is the probability of two wordsappearing and p(x) is the probability of a single wordappearing We use specific examples to explain this 0ereare three dispatching behavior words ldquoProvincial Dis-patchingrdquo ldquoDispatcher on dutyrdquo and ldquoProvincial Dispatcheron dutyrdquo If the word frequency of ldquoProvincial Dispatchingrdquois 10 the word frequency of ldquoDispatcher on dutyrdquo is 20 andthe word frequency of ldquoProvincial Dispatcher on dutyrdquo is 5the total number of words is N and the total number ofdouble words is M then we have the following formula

PMI Provincial DispatchingDispatcher on duty

log25M

(10N)lowast(20N)

(2)

0e mutual information can reflect the relationshipbetween two words well 0e higher the mutual informationvalue is the higher the correlation between X and Y is themore likely X and y are to form phrases On the contrary thelower the mutual information value is the lower the cor-relation between X and Y is the more likely there is a phraseboundary between X and y

Knowledge graph ofpower dispatching behavior

BILSTM‐CRF model

Entity recognition

Entity and relationship data

Power dispatching texts

Natural language processingsuch as text segmentation

Domainphrase

Manual annotation

Phrase extraction algorithmbased on mutual informationand the left and right entropy

Labeledcorpus

Dispatchingbehavior entity

Relationshipdefinition

Relationshipextraction

Text processing

Relationship extraction

Figure 1 Technical roadmap for constructing a knowledge graph of power dispatching behavior

Scientific Programming 3

Mutual information indicates the relevance of the twowords Also we need to calculate the degree of freedom ofthe word 0e degree of freedom refers to the degree ofdiversity of adjacent words that appear on the left and rightsides of the word If the left and right sides of a candidateword are different words in different sentences the smallerthe connection between the word and other words thegreater the internal connection between the candidatewords that is the greater the possibility that the candidatewords have boundaries and are a single word

We use the left and right entropy to measure the degreeof freedom Entropy can describe information uncertaintyIn information theoretic learning correntropy has been awidely used nonlinear similarity measure method due to itsrobustness [17] 0e larger the left entropy and right entropyof a candidate the more uncertain the words that mayappear on the left and right sides of the candidate and thehigher the degree of freedom0e formula for calculating theleft and right entropy is as follows

EL(W) minus 1113944forallaisinA

P(aW | W) middot log2 P(aW | W)

ER(W) minus 1113944forallbisinB

P(Wb | W) middot log2 P(Wb | W)(3)

Taking the left entropy as an example suppose that theldquoDispatcher on dutyrdquo has several kinds of collocationsldquoNational Dispatcher on dutyrdquo ldquoProvince Dispatcher ondutyrdquo and ldquoTemporary Dispatcher on dutyrdquo then the leftentropy of the word ldquoDispatcher on dutyrdquo is as follows

minus EL(Dispatcher on duty)

P(N D |Dispatcher on duty)

middot log2 P(N D |Dispatcher on duty)

+ P(P D |Dispatcher on duty)

middot log2 P(P D |Duty Officer)

+ P(T D |Dispatcher on duty)

middot log2 P(T D |Dispatcher on duty)

(4)

0e final input result is the score of a series of words0ecalculation formula of the score is as follows

score PMI + min(left entropy right entropy) (5)

0ese scores are sorted from high to low We add the top100 words to the jieba word segmentation dictionary andthen perform word segmentation processing on the originalcorpus text to facilitate the manual labeling of the role ofwords in the later period

312 Manual Annotation Due to the fuzzy boundary ofChinese words and a large number of cross-nesting struc-tures in the grid dispatching entity the complexity of theidentification task increases Furthermore the data set in thispaper contains multiple categories of entities Consequentlyaccording to the word segmentation results obtained by theunsupervised phrase extraction method the word seg-mentation results need to be returned to the original corpus

after manual inspection 0en we use the BMESO labelingmechanism to convert it to the input format required by themodel and finally get a labeled Training data set 0e def-inition of the BMESO annotation model is shown in Table 1

0e labeling tool we use is YEDDA For named entitiesin the field of power dispatching we have summarized manydispatch documents and dispatch glossary classificationmethods 0en a number of collaborators form a team tocollaborate to review and determine the dispatching be-havior entities and finally they are classified as follows

(1) Scheduling mechanism (SM) including Chinarsquos fivemajor power generation groups regional powergeneration groups State Grid Corporation of ChinaRegional Power Grid Corporation managementorganizations and departments at all levels

(2) Scheduling personnel (SP) including leaders ofvarious organizations technical personnel at alllevels and dispatching personnel on duty at all levels

(3) Scheduling operation (SO) including but not limited todispatching operation related to the protection device

(4) Facilities (Fac) such as transformer bus line circuitbreaker switch knife gate protection device pri-mary equipment secondary equipment electricalequipment boiler equipment steam (water gas)turbine equipment power transmission equipmenttransmission equipment converter equipmentpower system chemical treatment and fueltransportation

(5) Management requirements (MR) including sched-uling management scope (equipment name)scheduling management mode and schedulinginstructions

(6) Electric power data (EPD) such as power-relateddocuments systems and operation tickets

(7) Scheduling condition (SC) the objective conditions forcertain dispatching under the power performance suchas the conditions for the power outage and powerstations or substations on both sides of the line

(8) Equipment state (ES) such as operation mainte-nance standby charging power transmissionpower failure and other equipment states

32 Entity Extraction In the past few years the rapid de-velopment of machine learning has attracted the attention ofmany researchers [18] In order to identify and extractknowledge entities this paper uses a Bidirectional LongShort-Term Memory (BiLSTM) model and ConditionalRandom Fields (CRF) model as a named entity recognitionmodel We use the annotated data of the annotated corpusabove for model training and extraction of knowledge en-tities in the field of power dispatching behavior0e BiLSTMmodel is composed of forward LSTM and backward LSTM0e LSTM model can memorize the long-term dependenceof sentences from front to back but it cannot encode in-formation from back to front Compared with a single LSTMmodel the BiLSTMmodel can obtain bidirectional semantic

4 Scientific Programming

dependence and obtain more comprehensive text infor-mation However the BiLSTM model does not guaranteethat the prediction results obtained at each output layer arecorrect and some prediction results that do not meet theconstraints of the training set may appear 0erefore theCRF model can be introduced to learn the constrainingrules thereby reducing the output of the model the prob-ability of an illegal sequence 0e annotated corpus con-structed above prepares for the building of an entityrecognition model At the same time annotated data is usedfor model training to identify entities with domainknowledge of power dispatching behavior

0e BiLSTM+CRF model is mainly composed of threelayers and the schematic diagram of the model is shown inFigure 2 0e first layer is the embedding layer 0e wordvector is trained by inputting the pretrained character vectorand word vector and the dictionary obtained in the previouscorpus labeling process is added to make the generated wordvector more capable of expressing semantics

0e second layer in the middle is the forward andbackward LSTM layer In order to make full use of wordmeaning and word order information the input sequence ofthe character vector and the word vector of the matchingdictionary are subjected to feature fusion through networkcalculation

0e BiLSTM layer automatically extracts sentence fea-tures uses the char embedding sequence (x1 x2 x3 xn)of each word in a sentence as the input of each time step ofBi-LSTM and then uses the hidden state sequence(h1rarr

h2rarr

h3rarr

hn

rarr) output by the forward LSTM and the

reverse LSTM (h1

larr h2

larr h3

larr hn

larr) 0e hidden state output

at each position is stitched by a position to obtain a completehidden state sequence (h1

rarr h2rarr

h3rarr

hn

rarr) isinRnlowastm

0e output of this layer is the score of each label of aword by selecting the highest label score as the label of theword

Finally the CRF layer is introduced for sentence-levelsequence annotation 0e parameter of the CRF layer is a(k + 2) times (k + 2) matrix A k is the number of labels in thelabel set and Aij represents the transfer score from the i-thlabel to the j-th label When labeling a location you can usethe label that has been labeled before0e reason for adding2 is to add a start state to the beginning of the sentence andan end state to the end of the sentence Adding the CRFlayer can consider the order between the labels of theoutput words of the Bi-LSTM layer adding some con-straints to the last predicted label to ensure that the pre-dicted label is legal

0is paper introduces the Dropout mechanism to pre-vent overfitting 0e Dropout mechanism prevents over-fitting by randomly deleting hidden neurons in the networkwith a certain probability0e neurons in the input layer andthe output layer of the network remain the same In this waythe hidden neurons deleted in each iteration cycle are dif-ferent which increases the randomness of the network andimproves the generalization ability of the network 0emodel code is shown in Algorithm 1

33 Relationship Extraction In order to mine the rela-tionship of power dispatching behavior this paper needs toanalyze the language characteristics of the relationship de-scription of power dispatching text Since the power dis-patching text is an unstructured natural language textwritten in Chinese it has the characteristics of the Chineselanguage grammar and the power field Its specific char-acteristics are as follows

0e sentence contains a large number of power domainentities In a sentence related to scheduling behavior theremay be three or more behavior subjects and objects at thesame time 0e relationship network formed by the rela-tionship between any two entities in the sentence is com-plicated However the entity-relationship category isrelatively straightforward and the entity relationships be-tween the restricted entity categories mostly belong to onecategory

Each sentence in the dispatching text corresponds to adispatch behavior and each segment corresponds to a typeof dispatch scenario with various types Understanding thedispatch statement requires professional knowledge ofelectricity and it is difficult for nonprofessionals to learn0e Chinese grammatical structure is more flexible andsophisticated than English with many grammatical phe-nomena such as condition sequence causality and passiveDifferent writers have different language habits and dif-ferent scheduling behaviors will also use different expres-sions At present there is a lack of available syntacticknowledge rule base in the field of electric power

0e dispatching text which is the basis of dispatchingbehavior is based on the summary of real-world dispatchingbehaviors 0e content is refined the data volume is

B-SP M-SP M-SP M-SP E-SP

P1

h1

P2 P3 P4 P5

h2 h3 h4 h5

h1 h2 h3 h4 h5

CRF layer

LSTM output

Backward LSTM

Forward LSTM

Embedding

Figure 2 Schematic diagram of the BiLSTM-CRF model

Table 1 0e definition of the BMESO annotation model

Annotation MeaningB 0e first word of the entityM 0e internal words of theE 0e suffix of the entityS Single entity wordsO Nonphysical constituent words

Scientific Programming 5

inadequate and there is a lack of labeled data 0e char-acteristics of multiple entities in the sentence make itchallenging to label entity-relationship data Machinelearning algorithms commonly used in the general fieldoften require large amounts of labeled data and cannot bedirectly applied to power dispatch texts

Based on the above characteristics we define the typesof power dispatching behavior relations as shown inTable 2 In the knowledge graph the edges representingthe relationship have directions and the relationshipedges in different directions may have different rela-tionship types

According to the above definition most of the two entitieshave only one type of relationship If two entities appear in ageneral sentence and their entity type meets the predefinedrelationship it can be considered that there is a predefinedrelationship between the two entities To extract the entityrelationship if there are multiple entities of the same type in asentence there may be a special relationship between theseentities such as a union When analyzing power dispatchingbehavior sentences words such as ldquocommonrdquo ldquoparallelrdquoldquoandrdquo and ldquoorrdquo are often used in the sentence to express theorder parallel and other relationships If there are relatedwords in the sentence that represent particular sentencepatterns such as juxtaposition negation and time it can bedetermined that the sentence has a special relationship and aparticular relationship type For the dispatcher and dispatchoperation entity the relationship type between the two typesof entities is judged according to the position characteristics ofthe entity in the sentence If the dispatcher entity is before thedispatch operation entity the relationship arrow is directed bythe dispatcher to the dispatch operation Otherwise the re-lationship arrow is determined by the dispatch0e operationis directed to the dispatcher 0erefore this paper sorts outand extracts the entity relations of power dispatchingbehavior

34 Knowledge Graph Construction and Retrieval Afterextracting power dispatching behavior entities and rela-tionships we use a graph database to store entity and at-tribute information and rely on entity relationships to

connect directed edges between entity nodes thereby con-structing a knowledge graph structure We use a graphdatabase query language to provide a retrieval method basedon knowledge graphs Neo4j database is one of the morepopular graph databases with good performance and afriendly user interface We use the Neo4j database as astorage database to construct a knowledge graph for powerdispatching and use the declarative graph query languageCypher provided by the Neo4j database for knowledge graphretrieval

4 Experiment

Based on the knowledge graph construction method pro-posed above this paper presents the experimental work oflabeling corpus construction knowledge entity extractionand knowledge graph construction of power dispatchingbehaviors with the power dispatch text data set In thissection we will detail the experimental design experimentaldetails and experimental results

41 Data Sets and Data Preprocessing In this paper wecrawled 29 documents related to power dispatching behaviorsuch as power grid dispatching procedures basic knowledgeof dispatching and disposal plans of dispatching failure0ese documents were written by professional power dis-patchers and these documents fully describe the powerdispatching business process dispatching requirements anddispatching behavior of dispatchers in the dispatchingprocess In this paper the above documents are used as theoriginal corpus for entity extraction and knowledge graphconstruction experiments In order to facilitate the follow-upwork we unify the document format remove the spaces andnumbers in the document and leave only character-typedata

42 Experiments and Result Analysis

421 Construction of Power Dispatching Behavior AnnotatedCorpus and Entity Extraction 0ere are a large number ofunlabeled entity vocabularies in the field of power grid

Input selfOutput Trained model(1) Initialize the model(2) Define the Embedding layer(3) Add the Embedding layer to the model(4) Add forward LSTM to the modelunits 128 return_sequencesTrue(5) Add Dropout(6) Add backward LSTM to the modelunits 64 return_sequencesTrue(7) Add Dropout(8) Add TimeDistributed layer to the model(9) Define the CRF layer and Add the CRF layer to the model(10) Parameter status of each layer of the output model(11) Return model

ALGORITHM 1 How to build BiLSTM+CRF named entity recognition model

6 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

Mutual information indicates the relevance of the twowords Also we need to calculate the degree of freedom ofthe word 0e degree of freedom refers to the degree ofdiversity of adjacent words that appear on the left and rightsides of the word If the left and right sides of a candidateword are different words in different sentences the smallerthe connection between the word and other words thegreater the internal connection between the candidatewords that is the greater the possibility that the candidatewords have boundaries and are a single word

We use the left and right entropy to measure the degreeof freedom Entropy can describe information uncertaintyIn information theoretic learning correntropy has been awidely used nonlinear similarity measure method due to itsrobustness [17] 0e larger the left entropy and right entropyof a candidate the more uncertain the words that mayappear on the left and right sides of the candidate and thehigher the degree of freedom0e formula for calculating theleft and right entropy is as follows

EL(W) minus 1113944forallaisinA

P(aW | W) middot log2 P(aW | W)

ER(W) minus 1113944forallbisinB

P(Wb | W) middot log2 P(Wb | W)(3)

Taking the left entropy as an example suppose that theldquoDispatcher on dutyrdquo has several kinds of collocationsldquoNational Dispatcher on dutyrdquo ldquoProvince Dispatcher ondutyrdquo and ldquoTemporary Dispatcher on dutyrdquo then the leftentropy of the word ldquoDispatcher on dutyrdquo is as follows

minus EL(Dispatcher on duty)

P(N D |Dispatcher on duty)

middot log2 P(N D |Dispatcher on duty)

+ P(P D |Dispatcher on duty)

middot log2 P(P D |Duty Officer)

+ P(T D |Dispatcher on duty)

middot log2 P(T D |Dispatcher on duty)

(4)

0e final input result is the score of a series of words0ecalculation formula of the score is as follows

score PMI + min(left entropy right entropy) (5)

0ese scores are sorted from high to low We add the top100 words to the jieba word segmentation dictionary andthen perform word segmentation processing on the originalcorpus text to facilitate the manual labeling of the role ofwords in the later period

312 Manual Annotation Due to the fuzzy boundary ofChinese words and a large number of cross-nesting struc-tures in the grid dispatching entity the complexity of theidentification task increases Furthermore the data set in thispaper contains multiple categories of entities Consequentlyaccording to the word segmentation results obtained by theunsupervised phrase extraction method the word seg-mentation results need to be returned to the original corpus

after manual inspection 0en we use the BMESO labelingmechanism to convert it to the input format required by themodel and finally get a labeled Training data set 0e def-inition of the BMESO annotation model is shown in Table 1

0e labeling tool we use is YEDDA For named entitiesin the field of power dispatching we have summarized manydispatch documents and dispatch glossary classificationmethods 0en a number of collaborators form a team tocollaborate to review and determine the dispatching be-havior entities and finally they are classified as follows

(1) Scheduling mechanism (SM) including Chinarsquos fivemajor power generation groups regional powergeneration groups State Grid Corporation of ChinaRegional Power Grid Corporation managementorganizations and departments at all levels

(2) Scheduling personnel (SP) including leaders ofvarious organizations technical personnel at alllevels and dispatching personnel on duty at all levels

(3) Scheduling operation (SO) including but not limited todispatching operation related to the protection device

(4) Facilities (Fac) such as transformer bus line circuitbreaker switch knife gate protection device pri-mary equipment secondary equipment electricalequipment boiler equipment steam (water gas)turbine equipment power transmission equipmenttransmission equipment converter equipmentpower system chemical treatment and fueltransportation

(5) Management requirements (MR) including sched-uling management scope (equipment name)scheduling management mode and schedulinginstructions

(6) Electric power data (EPD) such as power-relateddocuments systems and operation tickets

(7) Scheduling condition (SC) the objective conditions forcertain dispatching under the power performance suchas the conditions for the power outage and powerstations or substations on both sides of the line

(8) Equipment state (ES) such as operation mainte-nance standby charging power transmissionpower failure and other equipment states

32 Entity Extraction In the past few years the rapid de-velopment of machine learning has attracted the attention ofmany researchers [18] In order to identify and extractknowledge entities this paper uses a Bidirectional LongShort-Term Memory (BiLSTM) model and ConditionalRandom Fields (CRF) model as a named entity recognitionmodel We use the annotated data of the annotated corpusabove for model training and extraction of knowledge en-tities in the field of power dispatching behavior0e BiLSTMmodel is composed of forward LSTM and backward LSTM0e LSTM model can memorize the long-term dependenceof sentences from front to back but it cannot encode in-formation from back to front Compared with a single LSTMmodel the BiLSTMmodel can obtain bidirectional semantic

4 Scientific Programming

dependence and obtain more comprehensive text infor-mation However the BiLSTM model does not guaranteethat the prediction results obtained at each output layer arecorrect and some prediction results that do not meet theconstraints of the training set may appear 0erefore theCRF model can be introduced to learn the constrainingrules thereby reducing the output of the model the prob-ability of an illegal sequence 0e annotated corpus con-structed above prepares for the building of an entityrecognition model At the same time annotated data is usedfor model training to identify entities with domainknowledge of power dispatching behavior

0e BiLSTM+CRF model is mainly composed of threelayers and the schematic diagram of the model is shown inFigure 2 0e first layer is the embedding layer 0e wordvector is trained by inputting the pretrained character vectorand word vector and the dictionary obtained in the previouscorpus labeling process is added to make the generated wordvector more capable of expressing semantics

0e second layer in the middle is the forward andbackward LSTM layer In order to make full use of wordmeaning and word order information the input sequence ofthe character vector and the word vector of the matchingdictionary are subjected to feature fusion through networkcalculation

0e BiLSTM layer automatically extracts sentence fea-tures uses the char embedding sequence (x1 x2 x3 xn)of each word in a sentence as the input of each time step ofBi-LSTM and then uses the hidden state sequence(h1rarr

h2rarr

h3rarr

hn

rarr) output by the forward LSTM and the

reverse LSTM (h1

larr h2

larr h3

larr hn

larr) 0e hidden state output

at each position is stitched by a position to obtain a completehidden state sequence (h1

rarr h2rarr

h3rarr

hn

rarr) isinRnlowastm

0e output of this layer is the score of each label of aword by selecting the highest label score as the label of theword

Finally the CRF layer is introduced for sentence-levelsequence annotation 0e parameter of the CRF layer is a(k + 2) times (k + 2) matrix A k is the number of labels in thelabel set and Aij represents the transfer score from the i-thlabel to the j-th label When labeling a location you can usethe label that has been labeled before0e reason for adding2 is to add a start state to the beginning of the sentence andan end state to the end of the sentence Adding the CRFlayer can consider the order between the labels of theoutput words of the Bi-LSTM layer adding some con-straints to the last predicted label to ensure that the pre-dicted label is legal

0is paper introduces the Dropout mechanism to pre-vent overfitting 0e Dropout mechanism prevents over-fitting by randomly deleting hidden neurons in the networkwith a certain probability0e neurons in the input layer andthe output layer of the network remain the same In this waythe hidden neurons deleted in each iteration cycle are dif-ferent which increases the randomness of the network andimproves the generalization ability of the network 0emodel code is shown in Algorithm 1

33 Relationship Extraction In order to mine the rela-tionship of power dispatching behavior this paper needs toanalyze the language characteristics of the relationship de-scription of power dispatching text Since the power dis-patching text is an unstructured natural language textwritten in Chinese it has the characteristics of the Chineselanguage grammar and the power field Its specific char-acteristics are as follows

0e sentence contains a large number of power domainentities In a sentence related to scheduling behavior theremay be three or more behavior subjects and objects at thesame time 0e relationship network formed by the rela-tionship between any two entities in the sentence is com-plicated However the entity-relationship category isrelatively straightforward and the entity relationships be-tween the restricted entity categories mostly belong to onecategory

Each sentence in the dispatching text corresponds to adispatch behavior and each segment corresponds to a typeof dispatch scenario with various types Understanding thedispatch statement requires professional knowledge ofelectricity and it is difficult for nonprofessionals to learn0e Chinese grammatical structure is more flexible andsophisticated than English with many grammatical phe-nomena such as condition sequence causality and passiveDifferent writers have different language habits and dif-ferent scheduling behaviors will also use different expres-sions At present there is a lack of available syntacticknowledge rule base in the field of electric power

0e dispatching text which is the basis of dispatchingbehavior is based on the summary of real-world dispatchingbehaviors 0e content is refined the data volume is

B-SP M-SP M-SP M-SP E-SP

P1

h1

P2 P3 P4 P5

h2 h3 h4 h5

h1 h2 h3 h4 h5

CRF layer

LSTM output

Backward LSTM

Forward LSTM

Embedding

Figure 2 Schematic diagram of the BiLSTM-CRF model

Table 1 0e definition of the BMESO annotation model

Annotation MeaningB 0e first word of the entityM 0e internal words of theE 0e suffix of the entityS Single entity wordsO Nonphysical constituent words

Scientific Programming 5

inadequate and there is a lack of labeled data 0e char-acteristics of multiple entities in the sentence make itchallenging to label entity-relationship data Machinelearning algorithms commonly used in the general fieldoften require large amounts of labeled data and cannot bedirectly applied to power dispatch texts

Based on the above characteristics we define the typesof power dispatching behavior relations as shown inTable 2 In the knowledge graph the edges representingthe relationship have directions and the relationshipedges in different directions may have different rela-tionship types

According to the above definition most of the two entitieshave only one type of relationship If two entities appear in ageneral sentence and their entity type meets the predefinedrelationship it can be considered that there is a predefinedrelationship between the two entities To extract the entityrelationship if there are multiple entities of the same type in asentence there may be a special relationship between theseentities such as a union When analyzing power dispatchingbehavior sentences words such as ldquocommonrdquo ldquoparallelrdquoldquoandrdquo and ldquoorrdquo are often used in the sentence to express theorder parallel and other relationships If there are relatedwords in the sentence that represent particular sentencepatterns such as juxtaposition negation and time it can bedetermined that the sentence has a special relationship and aparticular relationship type For the dispatcher and dispatchoperation entity the relationship type between the two typesof entities is judged according to the position characteristics ofthe entity in the sentence If the dispatcher entity is before thedispatch operation entity the relationship arrow is directed bythe dispatcher to the dispatch operation Otherwise the re-lationship arrow is determined by the dispatch0e operationis directed to the dispatcher 0erefore this paper sorts outand extracts the entity relations of power dispatchingbehavior

34 Knowledge Graph Construction and Retrieval Afterextracting power dispatching behavior entities and rela-tionships we use a graph database to store entity and at-tribute information and rely on entity relationships to

connect directed edges between entity nodes thereby con-structing a knowledge graph structure We use a graphdatabase query language to provide a retrieval method basedon knowledge graphs Neo4j database is one of the morepopular graph databases with good performance and afriendly user interface We use the Neo4j database as astorage database to construct a knowledge graph for powerdispatching and use the declarative graph query languageCypher provided by the Neo4j database for knowledge graphretrieval

4 Experiment

Based on the knowledge graph construction method pro-posed above this paper presents the experimental work oflabeling corpus construction knowledge entity extractionand knowledge graph construction of power dispatchingbehaviors with the power dispatch text data set In thissection we will detail the experimental design experimentaldetails and experimental results

41 Data Sets and Data Preprocessing In this paper wecrawled 29 documents related to power dispatching behaviorsuch as power grid dispatching procedures basic knowledgeof dispatching and disposal plans of dispatching failure0ese documents were written by professional power dis-patchers and these documents fully describe the powerdispatching business process dispatching requirements anddispatching behavior of dispatchers in the dispatchingprocess In this paper the above documents are used as theoriginal corpus for entity extraction and knowledge graphconstruction experiments In order to facilitate the follow-upwork we unify the document format remove the spaces andnumbers in the document and leave only character-typedata

42 Experiments and Result Analysis

421 Construction of Power Dispatching Behavior AnnotatedCorpus and Entity Extraction 0ere are a large number ofunlabeled entity vocabularies in the field of power grid

Input selfOutput Trained model(1) Initialize the model(2) Define the Embedding layer(3) Add the Embedding layer to the model(4) Add forward LSTM to the modelunits 128 return_sequencesTrue(5) Add Dropout(6) Add backward LSTM to the modelunits 64 return_sequencesTrue(7) Add Dropout(8) Add TimeDistributed layer to the model(9) Define the CRF layer and Add the CRF layer to the model(10) Parameter status of each layer of the output model(11) Return model

ALGORITHM 1 How to build BiLSTM+CRF named entity recognition model

6 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

dependence and obtain more comprehensive text infor-mation However the BiLSTM model does not guaranteethat the prediction results obtained at each output layer arecorrect and some prediction results that do not meet theconstraints of the training set may appear 0erefore theCRF model can be introduced to learn the constrainingrules thereby reducing the output of the model the prob-ability of an illegal sequence 0e annotated corpus con-structed above prepares for the building of an entityrecognition model At the same time annotated data is usedfor model training to identify entities with domainknowledge of power dispatching behavior

0e BiLSTM+CRF model is mainly composed of threelayers and the schematic diagram of the model is shown inFigure 2 0e first layer is the embedding layer 0e wordvector is trained by inputting the pretrained character vectorand word vector and the dictionary obtained in the previouscorpus labeling process is added to make the generated wordvector more capable of expressing semantics

0e second layer in the middle is the forward andbackward LSTM layer In order to make full use of wordmeaning and word order information the input sequence ofthe character vector and the word vector of the matchingdictionary are subjected to feature fusion through networkcalculation

0e BiLSTM layer automatically extracts sentence fea-tures uses the char embedding sequence (x1 x2 x3 xn)of each word in a sentence as the input of each time step ofBi-LSTM and then uses the hidden state sequence(h1rarr

h2rarr

h3rarr

hn

rarr) output by the forward LSTM and the

reverse LSTM (h1

larr h2

larr h3

larr hn

larr) 0e hidden state output

at each position is stitched by a position to obtain a completehidden state sequence (h1

rarr h2rarr

h3rarr

hn

rarr) isinRnlowastm

0e output of this layer is the score of each label of aword by selecting the highest label score as the label of theword

Finally the CRF layer is introduced for sentence-levelsequence annotation 0e parameter of the CRF layer is a(k + 2) times (k + 2) matrix A k is the number of labels in thelabel set and Aij represents the transfer score from the i-thlabel to the j-th label When labeling a location you can usethe label that has been labeled before0e reason for adding2 is to add a start state to the beginning of the sentence andan end state to the end of the sentence Adding the CRFlayer can consider the order between the labels of theoutput words of the Bi-LSTM layer adding some con-straints to the last predicted label to ensure that the pre-dicted label is legal

0is paper introduces the Dropout mechanism to pre-vent overfitting 0e Dropout mechanism prevents over-fitting by randomly deleting hidden neurons in the networkwith a certain probability0e neurons in the input layer andthe output layer of the network remain the same In this waythe hidden neurons deleted in each iteration cycle are dif-ferent which increases the randomness of the network andimproves the generalization ability of the network 0emodel code is shown in Algorithm 1

33 Relationship Extraction In order to mine the rela-tionship of power dispatching behavior this paper needs toanalyze the language characteristics of the relationship de-scription of power dispatching text Since the power dis-patching text is an unstructured natural language textwritten in Chinese it has the characteristics of the Chineselanguage grammar and the power field Its specific char-acteristics are as follows

0e sentence contains a large number of power domainentities In a sentence related to scheduling behavior theremay be three or more behavior subjects and objects at thesame time 0e relationship network formed by the rela-tionship between any two entities in the sentence is com-plicated However the entity-relationship category isrelatively straightforward and the entity relationships be-tween the restricted entity categories mostly belong to onecategory

Each sentence in the dispatching text corresponds to adispatch behavior and each segment corresponds to a typeof dispatch scenario with various types Understanding thedispatch statement requires professional knowledge ofelectricity and it is difficult for nonprofessionals to learn0e Chinese grammatical structure is more flexible andsophisticated than English with many grammatical phe-nomena such as condition sequence causality and passiveDifferent writers have different language habits and dif-ferent scheduling behaviors will also use different expres-sions At present there is a lack of available syntacticknowledge rule base in the field of electric power

0e dispatching text which is the basis of dispatchingbehavior is based on the summary of real-world dispatchingbehaviors 0e content is refined the data volume is

B-SP M-SP M-SP M-SP E-SP

P1

h1

P2 P3 P4 P5

h2 h3 h4 h5

h1 h2 h3 h4 h5

CRF layer

LSTM output

Backward LSTM

Forward LSTM

Embedding

Figure 2 Schematic diagram of the BiLSTM-CRF model

Table 1 0e definition of the BMESO annotation model

Annotation MeaningB 0e first word of the entityM 0e internal words of theE 0e suffix of the entityS Single entity wordsO Nonphysical constituent words

Scientific Programming 5

inadequate and there is a lack of labeled data 0e char-acteristics of multiple entities in the sentence make itchallenging to label entity-relationship data Machinelearning algorithms commonly used in the general fieldoften require large amounts of labeled data and cannot bedirectly applied to power dispatch texts

Based on the above characteristics we define the typesof power dispatching behavior relations as shown inTable 2 In the knowledge graph the edges representingthe relationship have directions and the relationshipedges in different directions may have different rela-tionship types

According to the above definition most of the two entitieshave only one type of relationship If two entities appear in ageneral sentence and their entity type meets the predefinedrelationship it can be considered that there is a predefinedrelationship between the two entities To extract the entityrelationship if there are multiple entities of the same type in asentence there may be a special relationship between theseentities such as a union When analyzing power dispatchingbehavior sentences words such as ldquocommonrdquo ldquoparallelrdquoldquoandrdquo and ldquoorrdquo are often used in the sentence to express theorder parallel and other relationships If there are relatedwords in the sentence that represent particular sentencepatterns such as juxtaposition negation and time it can bedetermined that the sentence has a special relationship and aparticular relationship type For the dispatcher and dispatchoperation entity the relationship type between the two typesof entities is judged according to the position characteristics ofthe entity in the sentence If the dispatcher entity is before thedispatch operation entity the relationship arrow is directed bythe dispatcher to the dispatch operation Otherwise the re-lationship arrow is determined by the dispatch0e operationis directed to the dispatcher 0erefore this paper sorts outand extracts the entity relations of power dispatchingbehavior

34 Knowledge Graph Construction and Retrieval Afterextracting power dispatching behavior entities and rela-tionships we use a graph database to store entity and at-tribute information and rely on entity relationships to

connect directed edges between entity nodes thereby con-structing a knowledge graph structure We use a graphdatabase query language to provide a retrieval method basedon knowledge graphs Neo4j database is one of the morepopular graph databases with good performance and afriendly user interface We use the Neo4j database as astorage database to construct a knowledge graph for powerdispatching and use the declarative graph query languageCypher provided by the Neo4j database for knowledge graphretrieval

4 Experiment

Based on the knowledge graph construction method pro-posed above this paper presents the experimental work oflabeling corpus construction knowledge entity extractionand knowledge graph construction of power dispatchingbehaviors with the power dispatch text data set In thissection we will detail the experimental design experimentaldetails and experimental results

41 Data Sets and Data Preprocessing In this paper wecrawled 29 documents related to power dispatching behaviorsuch as power grid dispatching procedures basic knowledgeof dispatching and disposal plans of dispatching failure0ese documents were written by professional power dis-patchers and these documents fully describe the powerdispatching business process dispatching requirements anddispatching behavior of dispatchers in the dispatchingprocess In this paper the above documents are used as theoriginal corpus for entity extraction and knowledge graphconstruction experiments In order to facilitate the follow-upwork we unify the document format remove the spaces andnumbers in the document and leave only character-typedata

42 Experiments and Result Analysis

421 Construction of Power Dispatching Behavior AnnotatedCorpus and Entity Extraction 0ere are a large number ofunlabeled entity vocabularies in the field of power grid

Input selfOutput Trained model(1) Initialize the model(2) Define the Embedding layer(3) Add the Embedding layer to the model(4) Add forward LSTM to the modelunits 128 return_sequencesTrue(5) Add Dropout(6) Add backward LSTM to the modelunits 64 return_sequencesTrue(7) Add Dropout(8) Add TimeDistributed layer to the model(9) Define the CRF layer and Add the CRF layer to the model(10) Parameter status of each layer of the output model(11) Return model

ALGORITHM 1 How to build BiLSTM+CRF named entity recognition model

6 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

inadequate and there is a lack of labeled data 0e char-acteristics of multiple entities in the sentence make itchallenging to label entity-relationship data Machinelearning algorithms commonly used in the general fieldoften require large amounts of labeled data and cannot bedirectly applied to power dispatch texts

Based on the above characteristics we define the typesof power dispatching behavior relations as shown inTable 2 In the knowledge graph the edges representingthe relationship have directions and the relationshipedges in different directions may have different rela-tionship types

According to the above definition most of the two entitieshave only one type of relationship If two entities appear in ageneral sentence and their entity type meets the predefinedrelationship it can be considered that there is a predefinedrelationship between the two entities To extract the entityrelationship if there are multiple entities of the same type in asentence there may be a special relationship between theseentities such as a union When analyzing power dispatchingbehavior sentences words such as ldquocommonrdquo ldquoparallelrdquoldquoandrdquo and ldquoorrdquo are often used in the sentence to express theorder parallel and other relationships If there are relatedwords in the sentence that represent particular sentencepatterns such as juxtaposition negation and time it can bedetermined that the sentence has a special relationship and aparticular relationship type For the dispatcher and dispatchoperation entity the relationship type between the two typesof entities is judged according to the position characteristics ofthe entity in the sentence If the dispatcher entity is before thedispatch operation entity the relationship arrow is directed bythe dispatcher to the dispatch operation Otherwise the re-lationship arrow is determined by the dispatch0e operationis directed to the dispatcher 0erefore this paper sorts outand extracts the entity relations of power dispatchingbehavior

34 Knowledge Graph Construction and Retrieval Afterextracting power dispatching behavior entities and rela-tionships we use a graph database to store entity and at-tribute information and rely on entity relationships to

connect directed edges between entity nodes thereby con-structing a knowledge graph structure We use a graphdatabase query language to provide a retrieval method basedon knowledge graphs Neo4j database is one of the morepopular graph databases with good performance and afriendly user interface We use the Neo4j database as astorage database to construct a knowledge graph for powerdispatching and use the declarative graph query languageCypher provided by the Neo4j database for knowledge graphretrieval

4 Experiment

Based on the knowledge graph construction method pro-posed above this paper presents the experimental work oflabeling corpus construction knowledge entity extractionand knowledge graph construction of power dispatchingbehaviors with the power dispatch text data set In thissection we will detail the experimental design experimentaldetails and experimental results

41 Data Sets and Data Preprocessing In this paper wecrawled 29 documents related to power dispatching behaviorsuch as power grid dispatching procedures basic knowledgeof dispatching and disposal plans of dispatching failure0ese documents were written by professional power dis-patchers and these documents fully describe the powerdispatching business process dispatching requirements anddispatching behavior of dispatchers in the dispatchingprocess In this paper the above documents are used as theoriginal corpus for entity extraction and knowledge graphconstruction experiments In order to facilitate the follow-upwork we unify the document format remove the spaces andnumbers in the document and leave only character-typedata

42 Experiments and Result Analysis

421 Construction of Power Dispatching Behavior AnnotatedCorpus and Entity Extraction 0ere are a large number ofunlabeled entity vocabularies in the field of power grid

Input selfOutput Trained model(1) Initialize the model(2) Define the Embedding layer(3) Add the Embedding layer to the model(4) Add forward LSTM to the modelunits 128 return_sequencesTrue(5) Add Dropout(6) Add backward LSTM to the modelunits 64 return_sequencesTrue(7) Add Dropout(8) Add TimeDistributed layer to the model(9) Define the CRF layer and Add the CRF layer to the model(10) Parameter status of each layer of the output model(11) Return model

ALGORITHM 1 How to build BiLSTM+CRF named entity recognition model

6 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

dispatching in the obtained power dispatch text data set Dueto professional domain issues these documents have nodistinct word boundaries 0en we use a phrase extractionalgorithm based on mutual information and left and rightentropy to extract domain words and use the extracteddomain words as a custom dictionary of Chinese wordssegmentation tool named ldquojiebardquo to assist in documentsegmentation As can be seen from the word segmentationresults in Figure 3 the use of the phrase extraction algorithmcan improve the quality of word segmentation and separatethe professional vocabulary in the power field such as HunanPower System and Relay Protection

According to the entity category of power dispatchingbehavior defined in this paper we complete the constructionof the labeled corpus of power dispatching behavior bymanually labeling the corpus after word segmentationShown in Figure 3 we use the code to build the BiLSTM-CRF model using an annotated corpus as the training set torealize the entity recognition of text for power dispatchingbehavior 0e recognition effect of the final model is shownin Figure 4 0e entity extraction method in this paper canextract the entity vocabulary of power dispatching behaviorfrom the power dispatching sentence and classify the en-tities It can be seen that the entity extraction method in this

Table 2 Predefined types of power dispatching behavior relations

Entity pairs Relational typeScheduling mechanismmdashscheduling operation Scheduling actionScheduling personnelmdashscheduling operation Scheduling actionScheduling operationmdashscheduling Instruction objectPersonnel facilitiesmdashscheduling condition Running stateFacilitiesmdashscheduling operation Scheduling modeScheduling conditionmdashscheduling operation Scheduling conditionScheduling conditionmdashscheduling condition Scheduling conditionmdashandornotManagement requirementsmdashscheduling operation Scheduling requirementsElectric power datamdashscheduling operation Scheduling basisScheduling operationmdashscheduling operation Scheduling behaviormdashorderandor

e Hunan Electric Power System appoints relevant dispatchers to perform relevant operations on the relay protection devices

e Hunan Electric Power System appoints relevant dispatchers to perform Relevant operations on the relay protection devices

Aer word segmentation

Before word segmentation

Figure 3 Partial word segmentation results of power dispatch text

(Onduty dispatcher)(The Hunan Electric Power System)

(The relay protection devices)

E-FACB-FAC M-FAC M-FAC M-FAC M-FAC O O O O O O

B-SP M-SP M-SP M-SP E-SPB-SM M-SM M-SM M-SM OOOOOE-SMM-SMIdentification

results

Chinese characters

Identification results

Chinese characters

Figure 4 Partial recognition results of power dispatching behavior entities

Scientific Programming 7

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

paper can extract the entity vocabulary of power dispatchingbehavior from the power dispatching sentence and classifythe entities

422 Construction of Knowledge Graph of Power DispatchingBehavior According to the relationship extraction methodmentioned above we extract the entity-relationship of the

Inform

Generator with no-load line for zero-start boost

Isolatingswitch

Generator loses

excitation and demagnetization protection

refuses to operate

Automationprofessional

Report

Open the disconnect switch on

both sides of the circuit

breaker

Open the bus tie breaker

invert other

operationRequires

measures to eliminate anomalies

Open the power plant side circuit

breaker

Processing

Dissection

Quit running

Increase excitation

Disconnect the

generator and

reconnect the grid

Reduce active output

Restore excitation

Trip

Abnormal

Non-full phase operation

Bus tiebreaker

AVC device

Circuitbreaker

Dispatch automation

system

AGC system

Generator

No-load line

Generator dragged into synchronizati

on

Generator high power

factor operationGenerator

phase advance

The generator is out of step

due to interference

System voltage allowed

The loss of excitation of the generator

did not destabilize the system

Dispatcher on duty

Operating staff

Site operation

regulations

Approved by the dispatcher

on duty

SchedulingCondition

SchedulingConditionAND

SchedulingCondition

SchedulingConditionAND

SchedulingConditionAND

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingCondition

SchedulingRequirementSchedulingConditionNotSchedulingConditionNot

Notice

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingConditionSchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

SchedulingAction

SchedulingBasis

InstructionObject

SchedulingAction

InstructionObject

InstructionObject

SchedulingCondition

InstructionObject

SchedulingCondition

RunningState

SchedulingAction RelativeDevices

SchedulingAction

SchedulingActionOrder

RelativeDevices

SchedulingActionOrder

SchedulingActionOrder

RelativeDevices

RelativeDevices

Components on the bus where the

circuit breaker is located to another group of buses for

Figure 5 Partial knowledge graph of power dispatching behavior

Circuitbreaker Two-phase

trip

Open theremainingtwo phases

Trip phasecannotclose

Single-phasetrip

Open-phaseoperation

Close thetrip phase

Open thenontripping

phase

Open the circuit breaker

Operator incharge

RunningState

RunningState

RunningState

SchedulingCondition

SchedulingAction

SchedulingCondition

SchedulingBebaviorOrder

SchedulingCondition

SchedulingCondition

SchedulingAction

SchedulingAction

Figure 6 Power dispatching behavior graph under the circuit breaker non-full phase operation scenario

8 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

power dispatching behavior based on the power dispatchingtexts and entity recognition results and form triples with theentity pairs 0e graph database Neo4j is used to store thedata and construct a knowledge graph structure 0e resultof the knowledge graph construction of power dispatchingbehavior is shown in Figure 5

0e nodes of different colors represent entities of dif-ferent entity categories Entities are connected by directededges that represent relationships between entities to formthe graph structure of the knowledge graph 0e knowledgegraph can store knowledge information such as knowledgeentities and relationships It is easy to see that comparedwith other forms of databases such as original text and tables

knowledge graphs link discrete data and knowledge rep-resentation and knowledge storage are more intuitive andefficient without the need for intermediate data conversionand processing

0is paper adds a ldquoscheduling scenariordquo attribute to therelationship of the knowledge graph to facilitate queryingthe possible scheduling behavior in a certain schedulingscenario in the knowledge graph Taking the schedulingscenario of the ldquonon-full phase operation occurs duringcircuit breaker operationrdquo as an example we executed theCypher query language of the neo4j database to conduct thequery 0e specific query statement is as follows

matchp () minus [lowast reSta ldquonon minus full phase operation occurs during circuit breaker operationrdquo1113864 1113865] minus ()returnp (6)

According to the query statement to get the powerdispatching behavior knowledge in this scenario the queryresult is shown in Figure 6 It can be seen from a simpleretrieval example that the power dispatch behaviorknowledge graph constructed in this paper has both se-mantic information and relationship information which canretrieve richer information and return intuitive visualizationresults In addition to the example retrieval method theknowledge graph query method is very flexible and can bequeried based on entity node attributes relationship attri-butes path depth etc to obtain richer knowledge infor-mation In the face of complex power dispatching businessthe knowledge graph constructed in this paper will provideknowledge about the dispatching behavior of related busi-nesses and effectively help dispatchers to conduct powerdispatching

5 Conclusion

0is paper explores the construction method of knowledgegraph based on power dispatching behavior In order toobtain the annotated corpus a phrase extraction algorithmbased on mutual information and left and right entropy isused in this paper to annotate the corpus by which thecorpus is constructed semiautomatically Based on the bi-directional long and short time memory network andconditional random field model the entity is trained andidentified0e relations of entities are extracted according tothe text of power dispatching behavior to store the data andconstruct the knowledge graph of power dispatchingbehavior

According to the constructed knowledge graph we cansearch more efficiently the knowledge related to the powerdispatching behavior provide the underlying knowledgemodel for the dispatching automation system and furtherimprove the intelligence of the power dispatching 0ereare also some problems and threats in this paper 0e dataset we used is small and the diversity of knowledgecontent requires more knowledge data support In

addition due to the lack of updated data we cannot studythe update process of the knowledge graph and the re-lationship extraction method in this article depends ontext mode and rules In the future we will conduct furtherresearch and improvement on the existing problemscontinue to explore a more efficient and automated re-lationship extraction model and study a more effectiveconstruction method of knowledge graph based on powerdispatching

Data Availability

0e data set contains some books of Grid DispatchingRegulations published by STATE GRID Corporation ofChina and its subsidiaries such as ldquoDispatching Regulationof Hunan Power Gridrdquo for Hunan province of China

Conflicts of Interest

0e authors declare no conflicts of interest

Authorsrsquo Contributions

For this paper Shixiong Fan conceived and designed theresearch study Shixiong Fan Zhifang Liao Xingwei Liuand Ying Chen collected data Shixiong Fan Xingwei LiuZhifang Liao Ying Chen and Yiqi Zhao designed themethodology and experiment Shixiong Fan Xingwei LiuYing Chen Yiqi Zhao and Huimin Luo completed theexperiment Shixiong Fan and Haiwei Fan conducted ap-plication deployment Ying Chen Yiqi Zhao and HuiminLuo wrote and modified the initial paper Zhifang LiaoYing Chen and Huimin Luo revised the paper All authorshave read and agreed to the published version of themanuscript

Acknowledgments

0is work was supported in part by the Basic ProspectiveProject of SGCC (no 5442DZ180017) and in part by the

Scientific Programming 9

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming

Science and Technology Research Foundation of SGCC(5442DZ180024-I)

References

[1] Y Bi L Jiang X Wang and L Cui ldquoDesign and investigationon service-oriented architecture-based smart grid dispatchingand control Systemrdquo Automation of Electric Power Systemsvol 39 no 2 pp 92ndash99 2015

[2] X Li J Xu Z GuoW Ning and ZWang ldquoConstruction andapplication of knowledge graph of power dispatch automationsystemrdquo China Electric Power vol 52 no 2 pp 70ndash77 2019

[3] T Steiner ldquoAdding realtime coverage to the google knowledgegraphrdquoProceedings of the 11th International Semantic WebConference (ISWC 2012) Boston MA USA September 2012

[4] Z Liao Z Zeng Y Zhang and X Fan ldquoA data-driven gametheoretic strategy for developers in software crowdsourcing acase studyrdquo Applied Sciences vol 9 no 4 p 721 2019

[5] Z Liao Z Wu Y Li Y Zhang X Fan and J Wu ldquoCore-reviewer recommendation based on Pull Request topic modeland collaborator social networkrdquo Soft Computing vol 24no 8 pp 5683ndash5693 2020

[6] Z Liao B Zhao S Liu et al ldquoA prediction model of theproject life-span in open source software ecosystemrdquo MobileNetworks and Applications vol 24 no 4 pp 1382ndash1391 2019

[7] Z Liao L Deng X Fan et al ldquoEmpirical research on theevaluation model and method of sustainability of the opensource ecosystemrdquo Symmetry vol 10 no 12 p 747 2018

[8] N Wang ldquoCompany name identification in Chinese financialdomainrdquo Journal of Chinese Information Pro Cessing vol 16no 2 pp 1ndash6 2002

[9] X Luo Y Li W Wang X Ban J-H Wang and W Zhao ldquoArobust multilayer extreme learning machine using kernel risk-sensitive loss criterionrdquo International Journal of MachineLearning and Cybernetics vol 11 no 1 pp 197ndash216 Jan 2020

[10] G Lample ldquoNeural architectures for named entity recogni-tionrdquo 2016 httpsarxivorgabs160301360

[11] J P C Chiu and E Nichols ldquoNamed entity recognition withbidirectional LSTM-CNNsrdquo Transactions of the Associationfor Computational Linguistics vol 4 pp 357ndash370 2016

[12] X Han Y Zhang W Zhang and T Huang ldquoAn attention-based model using character composition of entities inChinese relation extractionrdquo Information vol 11 no 2 p 792020

[13] J Leng and P Jiang ldquoA deep learning approach for rela-tionship extraction from interaction context in socialmanufacturing paradigmrdquoKnowledge-Based Systems vol 100pp 188ndash199 2016

[14] Y Tang ldquoEnhancement of power equipment managementusing knowledge graphrdquo in Proceedings of the 2019 IEEEInnovative Smart Grid Technologies-Asia (ISGT Asia) IEEEChengdu China May 2019

[15] Y Yang ldquoMulti-source heterogeneous information fusion ofpower assets based on knowledge graphrdquoProceedings of the2019 IEEE International Conference on Service Operations andLogistics and Informatics (SOLI) IEEE Zhengzhou ChinaOctober 2019

[16] H-F Wang ldquoAn error recognition method for powerequipment defect records based on knowledge graph tech-nologyrdquo Frontiers of Information Technology amp ElectronicEngineering vol 20 no 11 pp 1564ndash1577 2019

[17] X Luo J Sun L Wang et al ldquoShort-term wind speedforecasting via stacked extreme learning machine with

generalized correntropyrdquo IEEE Transactions on IndustrialInformatics vol 14 no 11 pp 4963ndash4971 2018

[18] M Chen Y Li X Luo W Wang L Wang and W Zhao ldquoAnovel human activity recognition scheme for smart healthusing multilayer extreme learning machinerdquo IEEE Internet ofgtings Journal vol 6 no 2 pp 1410ndash1418 2019

10 Scientific Programming