Consensus Based Distributed Reinforcement Learning for Nonconvex Economic Power Dispatch in Microgrids

A common assumption for economic power dispatch (EPD) is a perfect knowledge of cost functions. However, this assumption can be violated in cases when it is too difficult to establish an accurate model of the generation unit. In this paper, we formulate the EPD problem in a unified notation, based on which various reinforcement learning techniques can be applied. Then, a consensus based distributed reinforcement learning (CBDRL) algorithm is developed to solve the EPD problem. The CBDRL algorithm is fully distributed in sense that it requires only local computation and communication, which will contribute to a micro-grid of higher scalability and robustness. Finally, the effectiveness and performance of the proposed algorithm is verified through case studies.