Microgrid planning in time-division coupling mode for local consumption of RES: A nested reinforcement learning-distributionally robust optimization

• A novel planning framework for the microgrid based on a reconfigurable time-division coupling mode is proposed. • A planning method of nested energy-supply structure search and capacity-operation optimization is proposed. • A DRL-based flexible load aggregation and grouping method is presented. • A two-stage DRO method based on C&CG algorithm is proposed to address high-resolution RES uncertainties. Faced with the insufficient capacity of grids to accommodate renewable energy sources (RESs), many regions in China are stepping up efforts to promote source-load-storage integration within the microgrid (MG), prioritizing RES’s local consumption rather than exporting surplus power to the grid. However, blindly expanding the capacity of energy storage system (ESS) to mitigate RES fluctuations may compromise the economic viability of green RES investments. To address this challenge, this paper proposes a novel planning framework for MGs with flexible loads, based on a reconfigurable time-division coupling mode. By leveraging a clever power distributor (CPD), the framework dynamically matches RESs with flexible loads, enabling direct RES utilization and intuitive visualization of power supply–demand interactions. To solve the discrete programming problem of MG planning under dynamically shifting power supply structures, this paper proposes an integrated planning strategy that nests energy-supply structure search and capacity-operation optimization. In the search layer, the CPD supply structure is determined using a deep reinforcement learning (DRL) algorithm, which achieves faster convergence compared to traditional genetic algorithms. Considering the high-resolution uncertainty in RES outputs, a two-stage distributionally robust optimization based on the C&CG algorithm is employed in the optimization layer. The optimization results are then fed back into the DRL as reward signals to guide subsequent actions. The case studies show the proposed planning method reduces the total costs by 6.7 % and the ESS capacity by 20.8 % compared to traditional coupling mode.