Spatiotemporal graph neural network-based multi-node coordinated scheduling optimization for renewable energy power systems

The large-scale integration of wind and photovoltaic generation turns power-system dispatch into a strongly coupled spatiotemporal decision problem. Conventional optimization models become computationally demanding as system scale and uncertainty grow, whereas existing data-driven dispatch models usually rely on fixed topologies and cannot ensure physically admissible decisions. To address these issues, this paper proposes a physics-constrained spatiotemporal graph cooperative dispatch network (PC-STGCD) for multi-node coordinated scheduling of renewable-dominated power systems. First, an adaptive coupling graph fuses impedance-based electrical distance with similarity learned from node embeddings, so that weather-driven correlations among electrically remote stations can be identified. Second, a dual-branch encoder combines multi-head graph attention with multi-scale gated dilated convolution to jointly capture spatial cooperation and renewable fluctuations over different horizons. Third, power balance, ramping, branch-capacity, and storage-energy constraints are embedded into training as differentiable penalties, and a warm-started projection layer maps the network output onto the linearized feasible region during inference. On the modified IEEE 118-bus system, PC-STGCD reduces the operating cost by 6.94% and the renewable curtailment rate by 47.9% relative to the strongest baseline, limits the constraint violation rate to 0.43%, and produces a dispatch decision within 0.36 s; these improvements are statistically significant over five independent runs and persist under N−1 and N−2 contingencies without retraining. Validation on 1 year of measured data from a 236-node provincial grid yields an average cost saving of 5.84% and a 43.3% reduction in curtailment, demonstrating the potential of the proposed method for online cooperative dispatch of renewable-rich networks.