In new power systems, bidirectional DC-DC converters for energy storage serve as flexible hubs for accommodating high-penetration of renewable energy. However, under multiple uncertainties in microgrids, they face the risk of voltage instability. To address this issue, an observation-corrected active disturbance rejection control (ADRC) strategy assisted by the soft actor-critic (SAC) algorithm is proposed. First, two-dimensional disturbance information is incorporated into the compensation term to enable coordinated observation of disturbance state variables. Meanwhile, a hysteresis function is designed to compensate for the inherent shortcomings of the differential component, thereby accurately reconstructing the total disturbance. Subsequently, parameter tuning criteria are quantified, and the maximum entropy learning framework and stochastic policy exploration of the SAC algorithm are employed to achieve flexible matching of controller parameters under multiple disturbances, enabling the energy storage system to better perform peak shaving and valley filling. Finally, simulation comparisons under various operating conditions verify that the proposed strategy exhibits excellent dynamic performance under multiple internal and external uncertainties.