Designing a cross-regional emergency response network considering spatiotemporal evolution: A smart predict-then-optimise method

Emergency network design for future-oriented disaster management faces two key challenges: inter-regional disaster resilience and spatiotemporal heterogeneity. Traditional methods struggle with cross-regional collaborative response, heterogeneity modelling, and handling uncertain parameters. This study addresses this issue by proposing a two-stage cross-regional emergency response network design (SPO-2CERN), which integrates prediction and optimisation modules for smart decision-making. The first and second stages handle pre- and post-disaster emergency-management actions, respectively. First, we developed a deep learning model combining convolutional operations and multi-head self-attention, addressing the nonlinear mapping between heterogeneous disaster data and uncertain parameters. Second, we considered the robust design of emergency response networks with controllable risks and constructed a data-driven two-stage stochastic optimisation model. Subsequently, we developed a hybrid algorithm to solve the optimisation model efficiently. Empirical validation confirmed the effectiveness of the model in disaster prediction. Applying the SPO-2CERN framework to Gansu Province shows that its dynamic support-set update mechanism maintains two-stage cost optimality as risk levels vary and markedly improves decision stability. Across multiple risk strategies, it reduces redundant facility and inventory deployment costs by 6.8–14.9% relative to traditional schemes and lowers post-disaster transportation costs by 27.4–40.9%, while maintaining the optimality gap within 0.2% as the risk threshold increases. Furthermore, managerial implications of the framework were explored.