State Predictive Control of Modular SMES Magnet Based on Deep Reinforcement Learning

Abstract

Modular superconducting magnetic energy storage (M-SMES) system, which characterizes high reliability, flexibility, and strong scalability, can deal with the stability and economy of power sys-tem operation, large-scale renewable energy access, power quality and other issues. The thermal stability of M-SMES magnets is a key issue affecting its operation and coordinated control. In this paper, a deep reinforcement learning (DRL) based state predictive power allocation strategy, aiming at improving the reliability of M-SMES, is proposed. Firstly, the interaction between temperature, current, state of charge and other parameters is comprehensively analyzed, and a state database of SMES magnet is established. Then, the prediction model of magnet temperature rise is built. Based on the real-time state and the compensation demand from the grid side, a DRL algorithm is adopted to control each SMES module coordinately, which aims at maximizing the compensation capability of the M-SMES within a safe range. Finally, through a case study, the effectiveness of the proposed method is verified.