A Transparent Neural Network Solver for Optimal Power Flow Problem

Effectively solving the optimal power flow (OPF) problem is crucial for power system operations. However, the OPF represents a nonconvex and intractable problem, and using traditional optimization methods to solve it can result in sub-optimal or infeasible solutions. Recently, neural networks (NNs) have emerged as a promising technique for solving a large-scale OPF problem, but their black-box nature raises transparency and trust concerns in power system operations. To address this problem, this paper proposes a transparent NN for solving the OPF problem. The input of the proposed NN includes the operating conditions of a power grid, and its output represents the solution to the considered OPF problem. The proposed NN is transparent because its input-output relationship can be explicitly explained by a physical perspective, which is achieved by deriving the network weights from the Taylor expansion of the Karush–Kuhn–Tucke conditions of the OPF model. The results of the comprehensive numerical experiments demonstrate that the proposed transparent NN can achieve superior solution accuracy and reliability compared to the state-of-the-art NN-based methods while maintaining good solving efficiency and generalization ability.