MSPG-STGAT: A Spatiotemporal Prediction Framework for Plutonium Aerosol Dispersion in Explosion Scenarios over Twin-Hill Terrain Incorporating a Steady-State Wind Field Prior

摘要: Chemical explosion accidents in complex hilly terrain may cause the rapid dispersion of radioactive aerosols, creating a demand for fast and accurate prediction to support consequence assessment and emergency response. This study proposes a Multi-Scale Physics-Gated Spatio-Temporal Graph Attention Network (MSPG-STGAT), which integrates steady-state wind-field priors with multiscale spatiotemporal features to enable rapid prediction of the spatiotemporal evolution of concentration fields. Results show that MSPG-STGAT achieves a mean R² of 0.9041 on the test set, with a prediction time of approximately 5.76 s per scenario, and overall outperforms the baseline models. Ablation and error analyses demonstrate that the steady-state wind-field prior effectively improves model performance, with mean FAC3 values exceeding 0.92 across different spatial regions and temporal scales. Furthermore, generalization and robustness tests show that MSPG-STGAT maintains high predictive accuracy for an unseen double-hill terrain configuration and under wind-speed measurement errors. Its predictive performance is markedly superior to that of Hotspot, while its prediction time is approximately 2664 times shorter than that of CFD. These results demonstrate that MSPG-STGAT combines predictive accuracy, computational efficiency, and generalization capability, providing an effective approach to support rapid emergency decision-making for radioactive aerosol release scenarios caused by chemical explosions in complex hilly terrain.