Stable DC bus voltage is a critical prerequisite for achieving high utilization of new energy in DC microgrids. To address bus voltage fluctuations caused by uncertainties in source-load interactions, a flexible disturbance-rejection voltage regulation technique for energy-storage microgrids is proposed, based on deep reinforcement learning (DRL) with information entropy-based uncertainty quantification. First, the limitations of traditional disturbance compensation under uncertain energy interactions are analyzed. Second, information entropy evaluation index is used to quantify uncertainty, and DRL is introduced with the objective of transforming a disordered system into an ordered one, enabling flexible compensation and ensuring algorithmic convergence within the allowable voltage range. Furthermore, mechanism by which flexible compensation enhances system performance is examined from three perspectives: variable structure, variable parameters, and variable damping. Finally, simulation results verify that the proposed flexible compensation technology offers strong disturbance-refection capability under a wide range of uncertainties in source-load-storage interactions, significantly improving DC bus voltage stability.