大容量电池储能电站功率分配的决策变量多，且策略需考虑多个优化目标及能自动适应场景的不确定性。为此，提出了一种基于多智能体深度强化学习（multi-agent deep reinforcement learning，MADRL）的电池储能电站功率分配决策方法。首先，基于大容量电池储能电站结构及其功率分配特性构建基于MADRL的功率分配决策框架，每个储能单元设置一个功率分配智能体，多个智能体构成合作关系；然后，设计考虑储能电站有功功率损耗、荷电状态（state of charge，SOC）一致性和健康状态损失最小优化目标的功率分配智能体模型，采用深度确定性策略梯度（deep deterministic policy gradient，DDPG）算法去中心化训练各智能体网络参数，算法收敛后得到储能子系统充放电功率值。最后，算例验证了所提方法的有效性，能在有效提高储能子系统SOC均衡性的同时降低有功功率损耗、健康状态损失和充放电切换次数。

The decision variables for power allocation of large-capacity battery energy storage power station are numerous, and their strategies need to consider multiple optimization objectives and the uncertainty of automatically adapting to the scenario. Therefore, this paper proposes a power allocation decision-making method for battery energy storage power stations based on multi-agent deep reinforcement learning（MADRL）. Firstly, based on the structure and power allocation characteristics of large-capacity battery energy storage power stations, a power allocation decision framework based on MADRL is constructed. Each energy storage unit is equipped with a power allocation intelligent agent, and multiple agents form a cooperative relationship. Then, a multi-agent DRL model for power allocation is designed, considering the optimization objectives of active power loss, state of charge （SOC） consistency, and state of health loss of battery energy storage power stations. The deep deterministic policy gradient （DDPG） algorithm is used to decentralize the training of network parameters for each agent. After the algorithm converges, the charging and discharging power values of the energy storage subsystem are obtained. Finally, the effectiveness of the proposed method is verified by the example, which can effectively improve the SOC balance of the energy storage subsystem while reducing active power loss, state of health loss, and charging and discharging switching times.