基于物理信息强化学习的可再生能源实时最优潮流控制

The serious uncertainties from the extensive integration of renewable energy generations put forward a higher real-time requirement for power system dispatching. To provide economic and feasible generation operations in real-time, a physics-informed reinforcement learning (PIRL) method based on constrained reinforcement learning (CRL) for optimal power flow (OPF) is presented in this paper. In the proposed method, a physics-informed actor based on the power flow equations is designed to generate generation operations that satisfy the equality constraints of OPF. To specify inequality constraints in actor optimization, the policy gradient is augmented with the constraints to correct unfeasible generation operations. In particular, the cost functions related to inequality constraints can be directly calculated based on the output of the actor, which is more accurate than using networks to approximate in general CRL methods. The proposed method is tested on the IEEE 118-bus system, and the simulation result shows that the proposed method achieves a significant improvement in computation speed compared with the traditional interior point method while obtaining a similar generation cost.