Baselining PMU Data to Find Patterns and Anomalies CIGRE US National Committee 2015 Grid of the Future Symposium Brett amidan Jim Follum Kimberly Freeman Jeff Dagle Pacific Northwest National Laboratory November 15, 2018

“Big Picture” Objective Power grid related data (PMUs, State Estimators, Load, etc) Analytical Tool that provides: Real time analytics, monitoring the state of the grid Capability to look at historical trends and events Reliable predictions about the forthcoming state of the grid November 15, 2018

Pre-Processing Steps Read raw PMU data Develop and then use data quality filters to clean poor quality data Frequency Remove bad data Information about the data we are processing – 60 Hz PMU data for 50+ PMUs from BPA. We currently have processed 27+months of data (>8 TB). We are able to read, clean, and analyze 1 minute of data in under 45 seconds. 1 day of 60 Hz PMU data (54 PMUs) = 26 GB November 15, 2018 b.amidan@pnnl.gov

Feature Extraction (Data Signatures) Regression fits through the data calculate estimates of value, slope, curvature (acceleration), and noise. Can be calculated in the presence of missing or data quality flagged values. Summaries of these features are used in the analyses. November 15, 2018 b.amidan@pnnl.gov

Baselining Grid Behavior Univariate Approach Create a baseline of typical behavior for each individual variable Determine abnormal behavior based on the baseline Multivariate Approach Create a baseline across many (hundreds or even thousands) of variables Relationship between variables is considered when determining abnormal behavior Static Baselining Limits November 15, 2018

Univariate Baselining Example Date / Time Model – Time Series Based Model Day of Week Model Hours 0-23 Predicted Phase Angle Pair Value at Midnight Dynamic Baselining Limits (Calculated Daily) Phase Angle Difference Actual Value Initial Training Period November 15, 2018

Multivariate Baselining Baseline captures what normal behavior is expected to be Group similar behavior Time periods that group together indicate normal grid behavior Variables that group together indicate highly correlated variables and may be candidates for feature reduction Identify data that does not belong with the normal behavior Time period contains data that is unusual (possible abnormal grid behavior) Variable is unlike other variables, or something has happened to indicate a behavioral change in the variable November 15, 2018

Creating a Baseline – Unsupervised Learning Training Data: Historical PMU Data Baselining Learning Algorithm Class 1 Real Time PMU Data Model Class 2 Class 3 Class 4 Class 5 November 15, 2018

Identifying Data Driven Atypical Events Using multivariate statistical techniques to establish baselines of typical behavior, atypical moments in time can be discovered and the variables responsible can be identified. This slide shows how the atypicality score on the left increases due to atypical behavior in the system. The plots on the right show 2 different phase angle differences that were atypical during this same time period. November 15, 2018

Atypicality Detection Lightning Related Anomaly Atypicality Score Substation A Substation B November 15, 2018

Atypicality Detection Equipment Failure Related Anomaly Atypicality Score Other PMUs behaved similarly November 15, 2018 b.amidan@pnnl.gov

Phase Angle Pairs Clustering Unsupervised learning (clustering) used to determine which variables are most similar during Time Period A. Proximity on tree indicates similarity November 15, 2018 b.amidan@pnnl.gov

Phase Angle Pairs Clustering Time Period B (two months later) Phase Angle Pair #2 is no longer like Pair #1. Why? November 15, 2018 b.amidan@pnnl.gov

Baselining Learning Algorithm Supervised Learning Training Data: Historical PMU Data Class 1 Weather Baselining Learning Algorithm Class 2 Normal Class 3 Voltage Drop Class 4 Surge Class 5 Maintenance Labels Real Time PMU Data Predictive Model Weather Normal Voltage Drop Surge Maintenance November 15, 2018

Understanding Precursors to Inform Prediction Models Precursor Features (Signature) Precursor activity Inform Machine- Learning Model Create Classification to identify future precursors Known event November 15, 2018

Future Step – Using Supervised Learning to Predict Current State Possible Patterns Likelihood Event 1 0.75 Normal 4 0.15 Classification Based Prediction Model (Trained from Historical Data) Precursor 3 0.05 Extract Signature Event 7 0.04 NOTE: Only events and precursors with distinct data characteristics will be identifiable November 15, 2018

Conclusions Data driven anomalies can be identified using multivariate analyses techniques. Some of these anomalies correspond to actual events, but some do not. Understanding precursors can inform prediction models, allowing for probability based predictions of the near-term future grid behavior. November 15, 2018