Prescriptie Analytcs for Physical Systems Models
1
This Ph.D. is jointly funded by IT4BI-DC and AAU, Denmark Prescriptie Analytcs for Physical Systems Models References Olga Rybnytska 1 SolveDB DBMS 19 00 0 0 Supervisors: Torben Bach Pedersen 1 , Robert Wrembel 2 System Modeling and Predicton xtension 1. Gartner. htps://www.gartner.com/technology/home.jsp Accessed: 2017-12- 05. 2. Šikšnys, L., & Pedersen, T. B. (2016, July). Soliedb: Integratng optmizaton problem soliers into sql databases. In Proceedings of the 28th Internatonal Conference on Scientfc and Statstcal Database Management (p. 14). ACM. 3. Frazzeto, D., Nielsen, T. D., Pedersen, T. B., & Šikšnys, L. (2019). Prescriptie analytcs: a suriey of emerging trends and technologies. The VLDB Journal, 1-21. Aalborg University 1 Aalborg, Denmark Poznan University of Technology 2 Poznan, Poland Analysis/ Visualizato n (WolframAl pha) Optmizato n (CPLEX) Predicton (R) System Modeling (Modelica) Input Data Processing (PostgreS QL) Our Approach: Analytc Ascendancy Model [1] Example of a Typical Workfow (adapted from [3]): Analysis/ Visualizaton (partally by SolveDB+) Optmizaton (partally by SolveDB) Predicton Phase (pgFMU) System Modeling (pgFMU) Input Data Processing (SolveDB) Unifed PA tool (an integraton of SolieDB and upgraded pgFMU iersion to support Dynamic Optmizaton tasks) Extended SolieDB Architecture SolieDB query example [2] SolieDB [2] - a PostgreSQL-based DBMS with the natie support for in-DBMS optmizaton problem soliing. Proiides a set of built-in soliers for Linear Programming (LP)/Mixed Integer Programming (MIP), Global Optmizaton (GO) and domain-specifc problems. Allows SQL-based problem specifcaton and integrates solier into DBMS backend, therefore, making database-based problem solutons more easy and user- friendly. Motvaton: Complexity of model-driien analytcal tasks. Lack of Functonal Mockup Unit (FMI)-compliant model storage, simulaton and calibraton comprehensiie sofware support (within typical Python IDE), easy-to-use constructs and support within analytcal workfows. Limitatons when working with large amounts of data stored in non- iolatle memory. Step 1-5 executon tme comparison for 100 model instances of HP0, HP1, Classroom. For HP0, pgFMU+ is 2.43x faster, for HP1 – 2.85x faster, for Classroom – 3.4x faster RMSE comparison for 100 model instances of HP0, HP1 and Classroom. RMSE for all three models is almost identcal meaning pgFMU and pgFMU+ deliiers as accurate results as typical setup (Python) Prescriptve Analytcs: Motvaton and Challenges pgFMU - the frst DBMS extension to support simulaton, calibraton and ialidaton of Functonal Mockup Interface (FMI) – compliant physical systems dynamic models within a single DBMS en iironment. On aierage 2.9 tmes fas ter in comparison to the traditonal workfow, and 11.8 tmes less code lines. pgFMU is up to 12.5x faster in terms of deielopment tme for the arbitrary user-defned workfow. pgFMU system implementaton, usage scenario and results: 1. Create a DBMS model instance 2. Retrieie model iariables 3. Set model iariable ialues 4. Estmate parameters of the model 5. Simulate the model Steps 1-5 student workfow tme comparison for HP1 and Classroom (lef) and aierage user satsfacton with setups functonality comparison (right) Pseudocode for pgFMU User-Defned Functons (UDFs): fmu_create ()( Algorithm 1), fmu_simulate () (Algorithm 2), fmu_param_est () (Algorithm 3) Operaton Python pgFMU(+) 1. Create a DBMS model instance 4 1 2. Retrieie model iariables 12 1 3. Set model iariable ialues 4 1 4. Estmate parameters of the model 15 1 5. Simulate the model 24 1 Total 59 5 Lines of code comparison within typical setup (Python) and pgFMU(+) Running xample: User-defned arbitrary workfow to store, calibrate and simulate physical system model Linear tme-iniariant (LTI) state-space model of the heat pump heated room FMI physical systems model to test pgFMU upon PgFMU + - an optmizaton technique for estmatng parameters of multple model instances. The search space is reduced by using “inital guess” technique. Main idea: afer running Global Search + Local Search algorithms for one model instance, reuse the results for the remaining model instances, assuming the model instances are of the same structure and input tme series haie the same distributon. Pre-assessment questonnaire (1 - iery litle, 5 - iery much): Q1. How can you estmate your knowledge in energy systems and physical systems modeling? Q2. How familiar are you with model simulaton and model calibraton process? Q3. How familiar are you with model simulaton sofware(s)? Q4. How comfortable are you with using Python IDE? Q5. How comfortable are you with using SQL? Usability evaluaton : Conclusions:
Transcript of Prescriptie Analytcs for Physical Systems Models
This Ph.D. is jointly funded by IT4BI-DC and AAU, Denmark
Prescriptie Analytcs for Physical Systems Models
References
Olga Rybnytska1
SolveDB DBMS
1900
0
0
Supervisors: Torben Bach Pedersen1, Robert Wrembel2
System Modeling and Predicton xtension
1. Gartner. htps://www.gartner.com/technology/home.jsp Accessed: 2017-12-05.
2. Šikšnys, L., & Pedersen, T. B. (2016, July). Soliedb: Integratng optmizaton problem soliers into sql databases. In Proceedings of the 28th Internatonal Conference on Scientfc and Statstcal Database Management (p. 14). ACM.
3. Frazzeto, D., Nielsen, T. D., Pedersen, T. B., & Šikšnys, L. (2019). Prescriptie analytcs: a suriey of emerging trends and technologies. The VLDB Journal, 1-21.
Aalborg University 1
Aalborg, Denmark Poznan University of Technology2
Poznan, Poland
Analysis/ Visualizato
n (WolframAl
pha)
Optmizaton (CPLEX)
Predicton (R)
System Modeling
(Modelica)
Input Data Processing (PostgreS
QL)
Our Approach:
Analytc Ascendancy Model [1]
Example of a Typical Workfow (adapted from [3]):
Analysis/ Visualizaton (partally by SolveDB+)
Optmizaton (partally by
SolveDB)
Optmizaton (partally by
SolveDB)
Predicton Phase
(pgFMU)
Predicton Phase
(pgFMU)
System Modeling (pgFMU)
System Modeling (pgFMU)
Input Data Processing (SolveDB)
Input Data Processing (SolveDB)
Unifed PA tool (an integraton of SolieDB and upgraded pgFMU iersion to support Dynamic Optmizaton tasks)
Extended SolieDB Architecture
SolieDB query example [2]
SolieDB [2] - a PostgreSQL-based DBMS with the natie support for in-DBMS optmizaton problem soliing.
Proiides a set of built-in soliers for Linear Programming (LP)/Mixed Integer Programming (MIP), Global Optmizaton (GO) and domain-specifc problems.
Allows SQL-based problem specifcaton and integrates solier into DBMS backend, therefore, making database-based problem solutons more easy and user-friendly.
Motvaton: Complexity of model-driien analytcal tasks. Lack of Functonal Mockup Unit (FMI)-compliant model storage, simulaton
and calibraton comprehensiie sofware support (within typical Python IDE), easy-to-use constructs and support within analytcal workfows.
Limitatons when working with large amounts of data stored in non-iolatle memory.
Step 1-5 executon tme comparison for 100 model instances of HP0, HP1, Classroom. For HP0, pgFMU+ is 2.43x faster, for HP1 – 2.85x faster, for Classroom – 3.4x faster
RMSE comparison for 100 model instances of HP0, HP1 and Classroom. RMSE for all three models is almost identcal meaning pgFMU and pgFMU+ deliiers as accurate results as typical setup (Python)
Prescriptve Analytcs: Motvaton and Challenges
pgFMU - the frst DBMS extension to support simulaton, calibraton and ialidaton of Functonal Mockup Interface (FMI) – compliant physical systems dynamic models within a single DBMS eniironment.
On aierage 2.9 tmes faster in comparison to the traditonal workfow, and 11.8 tmes less code lines.
pgFMU is up to 12.5x faster in terms of deielopment tme for the arbitrary user-defned workfow.
pgFMU system implementaton, usage scenario and results:
1. Create a DBMS model instance
2. Retrieie model iariables
3. Set model iariable ialues
4. Estmate parameters of the model
5. Simulate the model
Steps 1-5 student workfow tme comparison for HP1 and Classroom (lef) and aierage user satsfacton with setups
functonality comparison (right)
Pseudocode for pgFMU User-Defned Functons (UDFs): fmu_create ()( Algorithm 1),fmu_simulate () (Algorithm 2),fmu_param_est () (Algorithm 3)
Operaton Python pgFMU(+)
1. Create a DBMS model instance 4 1
2. Retrieie model iariables 12 1
3. Set model iariable ialues 4 1
4. Estmate parameters of the model 15 1
5. Simulate the model 24 1
Total 59 5
Lines of code comparison within typical setup (Python) and pgFMU(+)
Running xample:
User-defned arbitrary workfow to store, calibrate and simulate physical system model
Linear tme-iniariant (LTI) state-space model of the heat pump heated room
FMI physical systems model to test pgFMU upon
PgFMU + - an optmizaton technique for estmatng parameters of multple model instances. The search space is reduced by using “inital guess” technique. Main idea: afer running Global Search + Local Search algorithms for one model instance, reuse the results for the remaining model instances, assuming the model instances are of the same structure and input tme series haie the same distributon.
Pre-assessment questonnaire (1 - iery litle, 5 - iery much):Q1. How can you estmate your knowledge in energy systems and physical systems modeling?Q2. How familiar are you with model simulaton and model calibraton process?Q3. How familiar are you with model simulaton sofware(s)?Q4. How comfortable are you with using Python IDE? Q5. How comfortable are you with using SQL?
Usability evaluaton :
Conclusions:
