Urban Flood Prediction Model Based on Transformer-LSTM-Sparrow Search Algorithm

Global climate change and accelerated urbanization have intensified extreme rainfall events, exacerbating urban flood risks. Although data-driven models have shown potential in urban flood prediction, the ability of single models to capture complex nonlinear relationships and their sensitivity to hyperparameters still limit prediction accuracy. To address these challenges, this study proposes an urban flood prediction model by integrating Transformer, Long Short-Term Memory (LSTM), and Sparrow Search Algorithm (SSA), combining Transformer’s global feature extraction with LSTM’s temporal modeling. The SSA was adopted to optimize hyperparameters for the Transformer-LSTM model. Dropout and early stopping techniques were adopted to mitigate overfitting. Applied to Zhengzhou city of Henan province, China, the model achieves a Nash-Sutcliffe Efficiency (NSE) of 0.971, indicating that the proposed model has high prediction performance for urban flooding. The experimental results demonstrate that the Transformer-LSTM-SSA model outperforms the standalone Transformer, LSTM, and Transformer-LSTM models by 12.9%, 10.1%, and 2.9% in NSE accuracy, respectively, while reducing MAE by 62.12%, 56.9%, and 34.21%, respectively, and MAPE by 21.69%, 22.2%, and 10.89%, respectively. Furthermore, the proposed model exhibits enhanced stability and superior generalization capability. The Transformer-LSTM-SSA model exhibits superior performance among the comparative methods, thereby demonstrating the model’s viability for providing a reliable solution for real-time flood prediction and early warning.