• physics-informed graph admittance attention network is designed to enhanced graph neural network’s perception of power grid topology. • A graph reinforcement learning-based solver is proposed for alternating current optimal power flow problem with N-1 topological contingencies. • Through the learning from demonstration technology, the training difficulties caused by the high-dimensional action space of large-scale power grids have been overcome, and outperforms traditional interior point solver. As a core non-convex NP-hard challenge in power system operation, the alternating current optimal power flow problem poses significant hurdles for traditional solvers in large-scale power grids. This study solves the alternating current optimal power flow problem with topological contingencies using graph reinforcement learning, which can effectively assist power system operators in making effective real-time decisions. To satisfy the strict physical constraints of power systems, the physical information is effectively encoded into an improved graph admittance attention network. Furthermore, learning from demonstration technology is introduced to develop a hybrid learning mechanism, which enables the agent to discover high-quality solution policies and accelerate the training convergence of the agent. Numerical experiments are conducted on the IEEE 30-bus, IEEE 118-bus systems and Illinois 200-bus benchmarktest system. The results demonstrate that the proposed solver outperforms conventional interior point solvers in both solution quality and computational efficiency.
Li et al. (Sat,) studied this question.