Online Scheduling of PV and Energy Storage System Based on Deep Reinforcement Learning

The Photovoltaic and Energy Storage System (PV-ES), as a typical microgrid, is increasingly become an important component of smart grid. Through effectively managing the energy operation of energy storage (ES), the PV-ES can effectively realize economic benefits, and provide power support for distribution networks (DN). However, the intermittency of PV power and the higher fluctuations of load power bring great uncertainty to the optimal scheduling of PV-ES. The optimality of traditional optimization methods for online scheduling can be insufficient because of the modeling complexity, strong conservatism, and heavy reliance on the prediction accuracy. To address these issues, this paper proposes an online ES scheduling algorithm for PV-ES based on deep reinforcement learning (DRL). Taking the system economic cost as the operation objective, DRL uses high dimensional PV-ES measurement data to directly drive the continuous action control of the energy storage, and can flexibly adapt to uncertain environments. Test results of a typical PV-ES show the superiority and effectiveness of this method.