Data Analytics and Machine Learning Approaches for Utility Companies

Modern Engineering and Technology Seminar 2018

Yannan Sun Data Scientist yannan.sun@oncor.com

Outline

• Introduction

• Statistical forecasting (Building energy use)

• Physics-based models (Parameter estimation for distribution systems)

• Statistical classification/anomaly detection (Building event detection)

• Oncor use cases using machine learning

About Me

• BS in Math at University of Sci. and Tech. of China

• MS in Statistics, PhD in Math at WSU

• Scientist/Senior Scientist at Pacific NW Natl. Lab

• Data Scientist at Oncor Electric Delivery

Oncor – By the Numbers

3.3M AMS Meters

121k T&D Circuit Line Miles

500k Switch Points

250k SCADA Points

1.5 B SCADA X-actions / month

60 TB New Data Records/month

250 TB Current EDW

Data storage

• Open energy delivery platform

• Enabling competitive retail,

generation, and open market in-home services

• Agnostic to generation

technologies or locations

• Focus on reliability and information services

• Need advanced data analytical methods

Oncor’s Analytics Platform

Data Size and Other Problems

Building Energy-Use Forecasting

using Regression Trees

Building Energy-Use Forecasting

• Volume – For one building the data is small, but if we want to mine the entire history of many buildings, the data size grows rapidly.

• Variety – numerical, categorical, ordinal, all kinds might be available from sensors.

• Velocity – hourly data with day ahead predictions, potentially not an issue.

• Veracity – weather reports, sensor malfunctions, lack of sensors

Building Energy-Use Forecasting

Inputs

Output

Holiday

Energy-Use Forecasting- Methods

• Several “Off-the-shelf” methods work well for prediction. – Regression tree, Gaussian Processes, Support Vector Machine (SVM),

(Deep) Neural Network

• Regression trees are fast for training and testing, so it is easy to use with a “rolling” window.

root

a>5 b>2

b<2

a<5

y = c11*a+c12*b

y = c21*a+c22*b

y = c31*a+c32*b

1-Day Ahead Forecasts – Regression Tree

• One month of initial training data • RMSE is about 12 • The model works well for all prediction windows, but the real dependency

is the weather forecast.

Normalized Error of Forecasts – 1 Day

Holidays

Training set too small

Parameter Estimation for Distribution

Systems

Parameter Estimation for Distribution Systems

• Volume – power injection, voltage, power flow, AMI measurements

• Variety – numerical data (multi-phase, multi-measurement)

• Velocity – SCADA data, 5-minute data generated by GridLAB-D (a power distribution system simulation tool)

• Veracity – bad data, missing data, parameter error, measurement noise

Parameter Estimation using the Kalman Filter

• Keeps track of the estimated state of the system and the variance or uncertainty of the estimate

• Many applications: dynamic positioning, satellite navigation systems, weather forecasting, power system state estimation, etc.

• State-vector augmentation with a Kalman filter

𝑥𝑘 = 𝑓(𝑥𝑘−1, 𝑢𝑘−1, 𝑤𝑘−1)

𝑧𝑘 = ℎ 𝑥𝑘 , 𝑣𝑘

where 𝑥 ≔ 𝑥𝑠 , 𝑥𝑝 with covariance 𝑃 ≔ 𝑃𝑠 00 𝑃𝑝

.

Parameter Estimation using the KF

Estimated parameter error with no measurement noise and single snapshot using residual analysis (rN) and the KF approach

Estimated parameter error with 1% measurement noise and single snapshot using residual analysis (rN) and the KF approach

Parameter Estimation using the KF

Estimated parameter error with 1% measurement noise and single time steps using residual analysis (rN) and the KF approach

Estimated parameter error with 1% measurement noise and 12 combined time steps using residual analysis (rN) and the KF approach

Parameter Estimation using the KF

𝜎(𝑧𝑘) 𝑡𝑐 = 1 𝑡𝑐 = 12 𝑡𝑐 = 24

RA KF RA KF RA KF

0.00% 0.0190 0.0178 0.0097 0.0094 0.0188 0.0188

0.50% 0.0174 0.0150 0.0091 0.0061 0.0180 0.0154

1.00% 0.0119 0.0180 0.0167 0.0190 0.0163 0.0192

1.50% 0.0026 0.0180 0.0048 0.0097 0.0138 0.0192

Mean of PE errors using RA and KF

𝜎(𝑧𝑘) 𝑡𝑐 = 1 𝑡𝑐 = 12 𝑡𝑐 = 24

RA KF RA KF RA KF

0.00% 0.0022 0.0043 0.0168 0.0174 0.0020 0.0041

0.50% 0.0626 0.0054 0.0259 0.0174 0.0132 0.0060

1.00% 0.1262 0.0042 0.0373 0.0039 0.0261 0.0035

1.50% 0.1922 0.0042 0.0595 0.0172 0.0393 0.0034

Standard deviation of PE errors using RA and KF

Building Event Detection using

Bayesian One-Class SVM

Building Event Detection

• Volume – for one building the data is small, but if we want to mine the entire history of many buildings, the data size grows rapidly.

• Variety – numerical, categorical, ordinal, all kinds might be available from sensors.

• Velocity – hourly data generated by EnergyPlus (a whole building energy simulation tool)

• Veracity – weather reports, sensor malfunctions, lack of sensors

Event/Anomaly Detection

• To predict binary outcomes (true/false) we could use supervised learning, this is called a binary classifier.

• The difficulty with this approach for faults is that it is very time-consuming to annotate the training data.

• Incorrect labels are also a problem. What is a fault? Are there unseen/unknown faults?

• To get around the labeling problem we adjust the problem as anomaly detection. This is an unsupervised classification problem.

Bayesian One-Class SVM

• The one-class SVM is used for novelty/outlier detection

• Construct the population density of a dataset and decide whether new points are – Normal: in high population – Abnormal: in low population

• The model is trained using past normal days. Predictions are made for the next week. – Unsupervised – Dynamic (time evolving)

Note that the frontier can be any level set, e.g. the light blue hourglass. Image from http://scikit-learn.org

Building Event Detection – Similarity Metrics

Radial basis function (RBF) Similarity

Euclidean Similarity

RBF wrong parameters

RBF wrong parameters

Building Event Detection– Data

Classification Results

Oncor Use Cases

Meter to Transformer Connectivity Error Detection

Transformer Load Forecast & Over Load Prediction

Data Analytics and Machine Learning Approaches for Utility Companies

Modern Engineering and Technology Seminar 2018

Yannan Sun Data Scientist yannan.sun@oncor.com