Heterogeneous graph attention-based deep reinforcement learning for optimal scheduling of flexible interconnected distribution substations

The spatial-temporal mismatch between generation and load is exacerbated by high distributed photovoltaic (PV) penetration in distribution service areas, causing power quality degradation and PV accommodation challenges. To tackle this issue, an end-to-end optimal scheduling method based on a heterogeneous graph attention network and deep reinforcement learning is proposed for intelligent coordinated control of flexible interconnected distribution service areas. First, the scheduling problem is formulated as a Markov decision process, and a heterogeneous graph structure is innovatively adopted to uniformly encode heterogeneous entities (transformers, PV units, energy storage systems, loads, tie lines) and their complex physical coupling relationships in the system. Second, a dual-layer graph neural network integrating node-level and type-level attention mechanisms is designed as the core perception and decision network of the deep reinforcement learning agent, enabling adaptive focus on critical system states and efficient learning of globally optimal scheduling policies. Finally, a simulation case based on real-world service area data is constructed for validation. Simulation results demonstrate that, compared with traditional linear programming, the proposed method reduces the total operating cost by 3.66%, PV curtailment by 52.86%, load-factor variance by 64.54%, and the maximum transformer load factor by 7.51%. Compared with ordinary deep deterministic policy gradient, the corresponding reductions are 0.88%, 35.34%, 46.90%, and 3.73%, respectively, while the training convergence time is shortened by 32.53%. These results demonstrate that heterogeneous graph-based state modeling improves both multi-objective scheduling performance and learning efficiency for flexible interconnected distribution substations.