Substation bus load forecasting and real-time regulation based on spatiotemporal graph neural networks and physics-prior multi-agent deep reinforcement learning

To address the problems of strong fluctuation, strong coupling, and difficulty in real-time regulation of substation bus load under the integration of distributed renewable energy and the participation of user-side adjustable resources, a bus load forecasting–real-time regulation closed-loop method based on a spatiotemporal graph neural network (STGNN) and physics-prior multi-agent twin delayed deep deterministic policy gradient (PI-MATD3) is proposed. First, a bus graph structure integrating electrical topology, load correlation, and voltage sensitivity is constructed, and multi-spatiotemporal-scale load forecasting and overload risk estimation are realized by using graph convolution, gated recurrence, and a scale attention mechanism. Second, the forecasting results are embedded into the PI-MATD3 framework, and source-load-storage collaborative regulation is achieved through topology-aware observation, action feasible-domain projection, and multi-scale control decomposition. Simulation results show that the average MAE, RMSE, and MAPE of STGNN under four types of scenarios are 0.035 MW, 0.0528 MW, and 1.49%, respectively, and the average R 2 is 0.9925. In high-risk scenarios, STGNN + PI-MATD3 reduces the voltage violation rate to 2.41%, increases the renewable energy utilization rate to 97.86%, and maintains good robustness under delay and noise disturbances. The study shows that the proposed method can improve the accuracy of bus-level short-term load forecasting, enhance the source-load-storage collaborative regulation capability, and take into account voltage security, resource dispatching cost, and online execution efficiency.