Bus voltage stability is a critical prerequisite for achieving high-level integration of new energy. To address the problem of bus voltage fluctuations in PV-storage DC microgrid with hybrid energy storage systems caused by source-load uncertainties, a self-driven optimal cascade active disturbance rejection control strategy integrated with the soft actor-critic (SAC) deep reinforcement learning algorithm is proposed. First, a cascaded extended state observer is designed to estimate and compensate for system uncertainties in real time, improving the accuracy of disturbance estimation. Then, a Markov decision model is established for the system, and a SAC agent, designed with a comprehensive evaluation of state rewards and information entropy, is integrated into the controller parameter optimization. By leveraging online learning and experience replay, the control parameters are autonomously and optimally tuned, further enhancing the system’s disturbance rejection capability and robustness. Finally, the performance of three control strategies under typical working conditions is compared by simulation experiment, validating the effectiveness and superiority of the proposed approach.