Abstract: With the large-scale integration of new energy into power grids, the intermittency and randomness of their output pose significant challenges to the stable operation of power systems. The rational capacity configuration of energy storage systems plays a crucial role in enhancing system reliability and promoting new energy consumption. This paper proposes an energy storage capacity configuration method based on Deep Reinforcement Learning (DRL) and Multi-Objective Particle Swarm Optimization (MOPSO). Firstly, it analyzes the operational characteristics of power systems and the access requirements of energy storage, constructs a system model incorporating wind power, solar power, and loads, and clarifies the configuration constraints and multi-objective optimization goals. Secondly, it utilizes DRL to learn energy storage operation strategies in complex environments; through the interaction between the DRL agent and the power grid environment, the agent outputs charging/discharging timing control recommendations. These recommendations are used as the particle initialization boundaries and search direction guidance for MOPSO. Combined with MOPSO's global search capability for energy storage capacity combinations, the method achieves a balance between the conflicts among economic, reliability, and new energy consumption objectives during the multi-objective optimization process. Case analysis based on an actual power grid shows that, compared with traditional methods, this approach can optimize the energy storage configuration scheme while satisfying system constraints, improve economic efficiency, reliability, and new energy consumption capacity. It provides a feasible technical means for energy storage planning under large-scale new energy integration and contributes to the low-carbon and efficient transformation of power systems.

Key words: deep reinforcement learning, multi-objective particle swarm optimization, energy storage system, global search capacity configuration