Residual Diffusion-Enhanced Spatio-Temporal Network for Human Mobility Forecasting in Disaster Scenarios

Forecasting human mobility during disasters is essential for emergency response but remains challenging because disasters cause abrupt, uncertain, and context-dependent deviations from regular mobility patterns. Deterministic models capture dominant spatio-temporal patterns but tend to underrepresent these deviations, whereas diffusion-based models characterize predictive uncertainty but may generate forecasts inconsistent with historical mobility dynamics. To overcome these limitations, we propose ResDiffNet, a residual diffusion-enhanced spatio-temporal network that decomposes mobility dynamics into a regular trend and a residual deviation. This decomposition enables a principled integration of deterministic and diffusion modeling: a deterministic base predictor first learns the stable mobility trend using a hypergraph convolutional network, and a residual conditional diffusion module (RCDM) subsequently models disaster-induced stochastic residuals through a conditional diffusion process. Guided by historical trend signals and disaster-related semantic information extracted from social media, the RCDM refines residual predictions in a context-aware manner, mitigating the excessive randomness of standard diffusion models. Experiments on two real-world disaster-related mobility datasets show that ResDiffNet reduces RMSE and MAE by 4.5–7.0% and 5.2–9.0%, respectively, relative to the strongest competing baseline for each task while achieving the lowest mean CRPS among the evaluated probabilistic methods. On Storm-New York (Outflow), CRPS performance is comparable to CoST, with no statistically significant advantage established. Moreover, integrating RCDM with four deterministic predictors consistently improves their forecasting accuracy, supporting its use as a plug-and-play residual refinement module.