Cross-regional hydro-meteorological cascade propagation modeling and emergency response optimization using temporal fusion transformer and spatio-temporal GNN
Year: 2026
Type: article
Abstract: Cross-regional hydro-meteorological disasters (e.g., floods, heavy rainfall-triggered cascades) often trigger cascading propagation effects with strong spatiotemporal dependencies due to geographical proximity, infrastructure interconnectivity, and socioeconomic coupling. Existing emergency response systems lack high-precision, interpretable predictive capabilities for such dynamic transmission processes, leading to delayed coordinated responses. To address this issue, this paper constructs a joint modeling framework integrating Temporal Fusion Transformer (TFT) and Spatio-Temporal GNN (ST-GNN). This framework inputs multi-source covariates from each region into the variable selection network of TFT. After gated residual processing, long-term dependencies are jointly modeled using multi-head self-attention and causal convolution, outputting regional temporal states. Simultaneously, ST-GNN constructs a task-adapted dynamic adjacency graph based on a predefined geographical adjacency matrix and adaptively learned edge weights. At each time step, neighbor nodes are aggregated and embedded, and node representations are updated via a time-gated mechanism. Both approaches align timestamps and concatenate features in each prediction step, generating risk intensity and prediction intervals through quantile regression. These are then used as constraint inputs for integer programming to solve resource scheduling schemes, forming a closed-loop decision. Comparative experiments show that this method achieves an average RMSE of 0.071 in the cascading propagation prediction of highly disturbed disaster distributions, a recall rate of 0.97, precision at 0.95, and an F1 score of 0.96 under extreme node network density, and an average response time of only 5.6 s under 500 nodes. As a proof-of-concept simulation framework, it achieves a synergistic improvement in prediction accuracy and decision-making efficiency, and provides quantifiable decision support for cross-regional emergency response, with retrospective case validation on a documented historical event.
Source: Discover Artificial Intelligence
Author: Longkai Chen
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $1,590
OpenAlex ID: W7225449341