Physics-Informed Multi-Agent Learning for Corrective Optimal Power Flow

Maintaining security is the priority for power system operation. The time window for the operator to take corrective action is quite limited post contingency. This work presents a novel physics-informed approach to the corrective optimal power flow (COPF) problem post contingency. COPF aims to determine a new optimal state post-contingency while meeting various physical laws and engineering limits. It is fundamentally important to find the corrective action when a contingency is detected. The traditional deep neural networks (DNNs)-based approach confronts challenges of explainability and solution accuracy. We introduce a physics-informed multiple DNNs to solve the problem. An affine matrix, containing physics information, is extracted from robust optimization for uncertainty management. Then it is fed to the multi-DNN model for network partitioning. We present a two-stage learning process with Lagrangian relaxation. Comprehensive case studies are conducted with IEEE 118- bus systems, IEEE 300- bus systems, and European 1354- bus systems. The proposed approach achieves an optimality gap of less than 0.1% while maintaining 99.9% constraint feasibility. Compared to the traditional optimization approach, the proposed method reduces the solution time by up to a factor of 414. The simulation results indicate that the proposed approach has improved accuracy and speed.