Graph Neural Network Enhanced Stochastic Optimization for Reliability Assessment of Composite Power Systems

Abstract

The increasing penetration of wind power intensifies the uncertainty and variability of power system operation. Monte Carlo simulation (MCS)-based reliability assessment of wind-integrated composite power systems usually requires repeated optimization over numerous operating states, thereby imposing a considerable computational burden. To address this issue, this paper proposes a graph convolutional network-enhanced stochastic optimization (GCN-SO) method. Through a dual-network architecture consisting of a classifier and a regressor, the proposed method jointly exploits power-grid topology and operating-state information to rapidly predict system load-shedding outcomes, thereby replacing repeated optimization under large numbers of operating states. Case studies on the RTS-79 system show that, under different wind power penetration levels, the reliability indices obtained by the proposed method are in close agreement with those obtained by the benchmark stochastic optimization method, while about a 25-fold computational speedup is achieved. Compared with existing neural network-based surrogate models, GCN-SO provides a better balance between prediction accuracy and computational efficiency. Component sensitivity analysis further shows that GCN-SO identifies critical components with significant impacts on system reliability in a manner consistent with the benchmark method. These results indicate that the proposed method can improve computational efficiency while maintaining reliability assessment accuracy, providing an effective tool for the reliability assessment of composite power systems with high renewable energy penetration.