Model Information-Aided Deep Learning for Corrective AC Power flow

This paper presents a novel model information-aided deep learning approach for corrective AC optimal power flow (OPF). The AC OPF power flow is a fundamental problem in the power system operation and planning. The corrective OPF is to find a new optimal state with flexible resources when contingency or extreme event occurs. AC OPF is regarded as nonconvex nonlinear optimization problem. Generally, there is no closed-form solution to the AC OPF. This work aims to use nonlinear neural network to approximate the corrective AC OPF solution with proper model information. The deep nonlinear neural network (DNNN) with decent depth and width is able to map numerous nonlinear functions. However, it often leads to slow training speed and poor accuracy to employ DNNN directly to solve OPF, especially for large size power systems. We propose to solve OPF with multi-networks with proper model information. An optimization approach is employed to feed special data to clustering for network partitioning. With multiple DNNNs, our approach achieves better performances in both training time and accuracy. The proposed techniques are validated and compared with other methods using multiple IEEE 118-bus system, IEEE 300-bus system, and European 1354-bus. The results of case studies demonstrate the effectiveness of the proposed approach.