Research on GNN-LSTM Model for Urban Waterlogging Depth Prediction

To address the challenges of low computational efficiency in traditional physical models and the difficulty of single neural networks in simultaneously capturing spatiotemporal features for urban waterlogging prediction, this paper proposes a hybrid GNN-LSTM model that integrates Graph Neural Networks (GNN) and Long Short-Term Memory networks (LSTM). The model aims to enhance the accuracy and timeliness of urban water depth prediction. Using Baotou City's Kundulun District as a case study, the research approach involves employing the Storm Water Management Model (SWMM) to simulate historical rainfall events and generate training data. A sub-catchment graph structure is constructed based on the actual drainage system topology, where GNN extracts spatial correlation features and LSTM learns temporal dynamics, achieving deep integration of spatiotemporal characteristics. Results show that the proposed model achieves an R² of 0.9454 and a Mean Absolute Error (MAE) of 0.0130 in short-term (t+1) predictions on the test set. The prediction accuracy is significantly superior to traditional methods, with a notable improvement in computational efficiency. In conclusion, the GNN-LSTM model effectively integrates the spatiotemporal evolution characteristics of urban waterlogging, providing reliable technical support for urban flood warning and emergency management, and demonstrating strong practical application value.