Phase Robust Optimization of PV-Energy Storage Microgrid Based on Deep Reinforcement Learning and Mixed Integer Constraint Model

The pronounced dependence of photovoltaic (PV) generation on meteorological conditions, coupled with substantial fluctuations in load demand, renders conventional deterministic optimization approaches inadequate. Addressing the need for robust multi-phase decision-making across temporal domains (e.g., day-ahead scheduling and real-time adjustment) and the coordinated optimization of continuous variables (such as energy storage charge/discharge rates) and discrete variables (such as unit commitment states), this research proposes a phased robust optimization strategy for PV-storage microgrids. This strategy integrates Deep Reinforcement Learning (DRL) with a Mixed-Integer Constrained Model (M-ICM). The methodology explicitly accounts for the coupling effects between irradiance intensity, temporal sequence efficiency, and the state-of-charge of energy storage systems. This ensures that the microgrid control system provides sufficient resilience mechanisms for dynamic energy allocation in practical applications, facilitating global optimization of microgrid energy utilization. The simulation results show significant improvements over conventional methods, which includes reduction in time-to-peak under dynamic balancing conditions, maintenance of lower output current-to-power ratios, and enhanced convergence speed of the neural network model.