Initial estimate of AC optimal power flow with graph neural networks

Optimal power flow (OPF) is a crucial task in power system management and control; accurate and time-efficient solutions for OPF are necessary to ensure cost-efficient and reliable power system operation. We introduce a novel solution to solving alternating current OPF (ACOPF), a nonlinear and nonconvex optimization problem, by combining the speed of a trained deep learning model with the accuracy of iterative solvers. The proposed framework uses a graph neural network (GNN) to exploit the graph structure of a power system in conjunction with proximal policy optimization, a deep reinforcement learning algorithm, to compute initial guesses for an interior point solver (IPS), providing a warm start, allowing the solver to converge in fewer iterations. Existing literature that explores warm start ACOPF solutions using machine learning choose to compute initial guesses that are trained to be feasible and cost-minimizing. Our approach trains the GNN-based reinforcement learning agent to produce an output that minimizes IPS convergence time by designing a reward function that is a function of the IPS convergence time. We evaluate the proposed framework using IEEE test case environments, using PyPower's IPS-based ACOPF solver and a GNN-based framework that computes ACOPF solutions directly as baselines, demonstrating significantly improved convergence times.