Data Driven Event Analysis for Cyber-Power System

Arman Ahmed

doi:10.7273/000005193

Power grid is prone to a diverse set of events, namely, power, cyber, and cyber-induced power events. Time-synchronized data from Phasor Measurement Units (PMUs) opens the possibility of developing data-driven efficient machine learning algorithms for monitoring such events. This dissertation proposes a novel set of supervised and unsupervised or semi-supervised machine learning models enabling real-time power and cyber event analysis using PMU data. This dissertation focuses on four research problems: (1) PMU data streams exhibit concept drifting properties, where the data distribution changes with operating conditions. Existing research fails to account for such a concept drifting nature in PMU data streams. This dissertation proposes a deep learning framework, “PMUNET”, that addresses the concept-drifting nature of PMU data streams. (2) Events in the power grid are spatial and temporal. Existing research either exploits the spatial or temporal nature but fails to exploit them jointly. In this dissertation, spatial-temporal models have been developed that jointly exploitthe spatial and temporal nature of PMU data, namely ‘Spatial-Temporal Graph Encoder and Decoder’ for transmission and ‘Spatial-Temporal Graph Autoencoder’ for the distribution system. (3) The power grid has increasingly become prone to cyberattacks. There is a need to develop an event analysis technique to monitor cyber-power events. A supervised ensemble decision tree algorithm has been developed for the transactive energy system. Also, an unsupervised deep learning autoencoder algorithm is developed for the transmission protection system. A spatial-temporal neural network has been developed for the distribution system. (4) PMU data in the power grid flows from the substation to the operation center hierarchically and can be processed in a distributed setting. Existing research on analyzing PMU data for control has followed a centralized approach where the data is sent to a lead computing agent. Sending data adds computational time and storage overload, leading to a delay in control. This dissertation proposes a federated inference framework that accounts for the distributed nature of PMU data for fast, proactive stability control. The proposed machine learning models for grid monitoring and control are validated using PMU measurement data from different IEEE test systems and real-world data, achieving promising results.

Data Driven Event Analysis for Cyber-Power System

Files and links (1)

Abstract

Metrics

Details