Abstract：In response to the intermittency and uncertainty of wind power generation，which result in low wind

power utilization and high electricity purchase costs in microgrids，an intelligent dispatch strategy based on a

residual-like soft actor-critic（R-SAC）algorithm was proposed. The scheduling problem was formulated as a

partially observable Markov decision process（PO-MDP），and short-term predictions of wind power output and

load variations were achieved by combining long short-term memory（LSTM）networks with the attention

mechanism. A residual-like network structure was then incorporated into the traditional soft actor-critic（SAC）

framework to construct an improved R-SAC algorithm，which enhanced the convergence speed and exploration

efficiency of the policy. Finally，simulation experiments based on data from an actual microgrid in the northwest

region validated the effectiveness and superiority of the proposed strategy.

      Key words：microgrid scheduling；wind power utilization；deep reinforcement learning（DRL）；state estimation；policy optimization