Yue Zhao / Stony Brook University

Abstract: Real-time power system monitoring is at heart an inference problem. In contrast to physical model based inference methods, we present a “Learning-to-Infer" framework where a) Information from the physical model is captured by the data generated with it, and b) powerful predictors are trained offline for accurate online inference. We present case studies on two fundamentally hard problems in power system monitoring, multi-line outage identification and voltage stability margin estimation, to illustrate the power of the Learning-to-Infer method. The achievement of two key properties by the method will be discussed: a) generalizability: the predictors achieve high performance on instances, not only unseen in, but also not similar to the training set, and b) scalability: with a moderate increase of the offline generated training data, the predictors achieve sustained high performance in larger systems.

Learning Materials: Talk Flyer Talk Slides

Interacting Materials:

1. Papers
   1. A Learning-to-Infer Method for Real-Time Power Grid Multi-Line Outage Identification
      
   2. LEARNING TO INFER: A NEW VARIATIONAL INFERENCE APPROACH FOR POWER GRID TOPOLOGY IDENTIFICATION
      
   3. Monitoring of Power System Topology in Real-Time
   4. Complex Power System Status Monitoring and Evaluation Using Big Data Platform and Machine Learning Algorithms: A Review and a Case Study
   5. A Learning-to-Infer Method for Real-Time Power System Monitoring(1)
   6. A Learning-to-Infer Method for Real-Time Power System Monitoring(2)
   7. Deep Reinforcement Learning for Power System Applications: An Overview
2. Videos
   1. Video Camera-based Structural Monitoring - Motion Magnification
      
   2. Toward an Integration of Deep Learning and Neuroscience
      
   3. POWER SYSTEM PERFORMANCE MONITORING
      
   4. Real-Time - Power System Monitoring and Simulation
      
   5. Monitor energy meters to feed The Smart Grid
      
3. Data