To address the challenges posed by the strong volatility of distributed energy resources and dual uncertainty in both generation and load within high-penetration distribution networks, this paper proposes a spatiotemporal load forecasting model integrating “Graph Neural Networks (GNN) + Transformer” alongside a shared-reward hierarchical reinforcement learning (HRL) energy management framework. Firstly, the power-distance composite weighted graph is constructed, and the geographical diffusion effect of node load is captured by Chebyshev graph convolution. Then, the Transformer encoder is connected to mine multi-time scale dynamic dependence, and the soft constraints of power balance and total load smoothing are embedded in the loss function to improve the physical interpretability of the forecast. The prediction results are output in a sequence of 15 min and 96 points, which are used to drive the HRL strategy: the upper agent formulates the rough scheduling of power purchase, unit start-stop and energy storage before the day to minimize the operating cost and reserve the adjustment margin for the lower layer; Based on ultra-short-term prediction, the lower agents fine-tune the energy storage and flexible load in real time, share coupling rewards to punish SOC overrun and voltage deviation, and realize "day-to-day-to-real-time" seamless cooperation. The experiment of IEEE 33-node high-permeability PV+ESS scenario shows that the proposed model reduces the MAE to 19.5 kW and 33.6 kW on typical days and extreme days respectively, which is 56.9% and 57.4% lower than that of ARIMA. The total cost of HRL scheme in 7 days is 12 810 yuan, the voltage exceeds the limit only 5 times, and the PV absorption rate is 96.8%, which is significantly better than the traditional rules and single-layer DDPG method, and provides a new accurate, robust and economical intelligent dispatching method for high permeability distribution networks.
Zhang et al. (Sun,) studied this question.