The use of machine learning and AI in analysis of historical and on-line synchrophasor data

Date: Tuesday, May 25                              Time:

Name of the organizer: Mladen Kezunovic, Regents Professor


Organisation: Texas A&M University, USA

Short biography of the chair: Mladen Kezunovic (S’77–M’80–SM’85–F’99– LF’17) received the Dipl. Ing. from University of Sarajevo, Sarajevo, Bosnia, and M.Sc. and Ph.D. degrees in electrical engineering from University of Kansas, Lawrence, KS, in 1974, 1977, and 1980, respectively. He has been with Texas A&M University, College Station, TX, USA since 1986, where he is currently Regents Professor, Eugene E. Webb Professor, and the Site Director of “Power Engineering Research Center” consortium. For over 25 years he has been the Principal Consultant of XpertPower Associates, a consulting firm specializing in power systems data analytics. His expertise is in protective relaying, automated power system disturbance analysis, computational intelligence, data analytics, and smart grids. He has authored over 600 papers, given over 120 seminars, invited lectures, and short courses, and consulted for over 50 companies worldwide. Dr. Kezunovic is a CIGRE Fellow, Honorary and Distinguished member. He is Registered Professional Engineer in Texas.

Abstract: The synchrophasor network in the USA has been expanding over the last decade and has reached over 3000 PMUs installed in the three interconnections: The West, East and ERCOT. The Transmission Operators and Independent System Operators have collected large amounts of historical PMU data. The challenges in searching through the data, detecting events of interest, an classifying them have been recognized. The Department of Energy has funded eight on-going projects to develop Artificial Intelligence (AI) and Machine Learning (ML) techniques to analyze and classify the events automatically. This panel will be focusing on sharing the experiences from four projects that have focused on different issues associated with the automation of synchrophasor data analytics: data wrangling, bad data detection and removal, reconstruction of missing data, data mislabeling, supervised, semi-supervised, unsupervised and transfer learning, deep learning, etc.

Panelist 1:

Name: Philip Hart

Organisation: GE Research, USA

Title of presentation: PMU-Based Data Analytics Using Digital Twin and Phasor Analytics Software

Abstract: Application of a powerful, industry-validated signature identification strategy to a large PMU dataset is described. The event signatures generated using the semi-supervised strategy are derived from an over-abundance of features calculated in a transparent manner, and can be efficiently applied to either historical or streaming PMU data. The signatures can be used to quantify the relative severity, location, and duration of each event, and show promise for integration into tools that provide enhanced power systems reliability, operational efficiency and resiliency.

Panelist 2:

Name: Bruno Leao

Organisation: Siemens Technology, USA

Title of presentation: MindSynchro: innovation and challenges in application of ML to PMU big data

Abstract: The presentation describes recently developed innovations associated to the application of ML methods to real world PMU big data. Flagship innovation comprises semi-supervised learning approaches for detection of relevant grid events which can be applied to a very broad PMU fleet. Other innovations and challenges are also discussed, especially concerning the relevance of adequately labeling PMU data for proper extraction of its value and approaches for overcoming existing limitations in such information.

Panelist 3:

Name: Zoran Obradovic

Organisation: Temple University – L.H. Carnell Professor, USA

Title of presentation: BDSmart: Automated Analysis of Large synchrophasor datasets

Abstract: We will discuss tradeoffs in machine learning and AI applications to automated PMU data analysis: preprocessing vs raw data, feature engineering vs inaccurate labels, computational implicity vs complexity, supervised vas unsupervised learning. We will then present some comparative results from different algorithms adopted by us to the problem at hand. Some lessons learned will be shared and discussed.

Panelist 4:

Name: Nanpeng Yu

Organisation: University of California, Riverside, USA

Title of presentation: Physics-Informed Machine Learning for Power System Event Detection and Identification with Synchrophasor Data

Abstract: Three recent breakthroughs in physics-informed synchrophasor data analytics will be reported in this panel presentation. First, we will talk about how to construct a graph Laplacian using off-line training data and detect power system events using graph signal processing techniques. Second, we will talk about how to leverage deep neural networks enhanced augmented by information loading and graph-based sorting to classify power system events based on PMU data.