Abstract: To address the challenges in the optimal scheduling of microgrid clusters, including difficulties in global coordination, complex local constraints, and high uncertainty in wind and photovoltaic power outputs, a hierarchical cooperative scheduling method based on an improved soft actor–critic (SAC) algorithm and an improved particle swarm optimization (PSO) algorithm is proposed. A two-layer optimization model is constructed, consisting of a microgrid cluster aggregator at the upper layer and multiple sub-microgrids at the lower layer. The upper layer is responsible for overall power coordination of the microgrid cluster, while the lower layer handles the autonomous optimization within each sub-microgrid. For the high-dimensional continuous decision-making problem at the upper layer, a variational autoencoder (VAE) is introduced for state compression, and the SAC algorithm is enhanced with a time-decay prioritized experience replay mechanism to improve learning efficiency and environmental adaptability. For the lower-layer optimization problem, which involves multiple constraints and nonlinearity, an improved PSO algorithm is designed, incorporating adaptive inertia weights, elite leader selection, and constraint repair mechanisms to enhance its ability to solve problems within complex feasible regions.