Objective To address the spatiotemporal dynamic evolution of regional disaster losses in sudden disaster management, as well as the resulting high uncertainty in post-disaster emergency demand and road segment capacity.

Method This paper proposes an integrated intelligent prediction and stochastic optimization framework for multi-stage cross-regional emergency response network design, namely the smart predict-then-optimize for multi-stage cross-regional emergency response network （SPO-MCERN）, to coordinate pre-disaster preparedness and post-disaster response decisions. First, to overcome the limitations of traditional statistical methods in capturing the complex spatiotemporal evolution of disaster losses, a hybrid deep learning prediction model, Conv-MHSA, is developed by integrating one-dimensional convolutional operations （1D-Conv） with a multi-head self-attention （MHSA） mechanism. Based on the constructed high-dimensional regional disaster feature set, the model employs an odd-even feature separation strategy for effective dimensionality reduction, thereby achieving high-accuracy prediction of regional disaster losses. Secondly, based on the prediction outputs, a dynamic assessment mechanism incorporating regional correction and seasonal response is designed to quantify post-disaster dynamic demand and road segment capacity. Finally, the SPO-MCERN model is reformulated using the sample average approximation （SAA） method, and a hybrid stochastic optimization algorithm combining Monte Carlo simulation and SAA is proposed.

Result Based on the historical data of earthquake disasters in China, Conv-MHSA demonstrates significant advantages in terms of prediction accuracy, robustness and overall performance. The case study of the emergency response network design in Gansu Province shows that the SPO-MCERN framework enables the network design and transportation costs to remain optimal under the conditions of temporal and spatial evolution.

Conclusion Compared with traditional approaches, the coordinated predict-then-optimize decision-making mechanism reduces redundant facility and material deployment costs by 21.6%, while correspondingly reducing post-disaster transportation and procurement costs by 25.9%. It can provide scientific support for cross-regional emergency decision-making in disaster scenarios.