Graph Learning-Based Power System Health Assessment Model

Abstract

As the power transmission system’s energy sources become increasingly diversified, the grid stability is experiencing increased fluctuations, thereby necessitating more frequent and near real-time monitoring by grid operators. The power system security has been monitored through real-time contingency analysis and dynamic security assessment framework, both of which are typically based on time-domain simulations or power flow calculations. Achieving higher accuracy in grid health level prediction often requires time-consuming simulation and analysis. To improve computational efficiency, this paper develops machine learning models with phasor measurement unit (PMU) data to monitor the power system health index, focusing on rotor angle stability and frequency stability. The proposed machine learning models accurately predict frequency and angle stability indicators, essential for evaluating grid health considering various contingencies, even when dealing with limited PMU deployment in transmission grids. The proposed framework leverages a physics-informed graph convolution network and graph attention network with ordinal encoders, which are benchmarked with multi-layer perceptron models. These models are trained on dataset derived from an augmented IEEE 118-bus system with different demand levels and fuel mix, including tailored dynamic generator models, generator controller models, and grid protection models. The numerical studies explored the performance of the proposed and baseline machine learning models under both full PMU coverage and various partial PMU coverage conditions, where different data imputation methods are used for substations without PMUs. The findings from this study offer valuable insights, such as machine learning model selection and critical PMU locations regarding power equipment, into the design of data-driven grid health index prediction models for power systems.