An interpretable hybrid deep learning model for flood forecasting based on Transformer and LSTM

Abstract

Study region: Flood formation involves complex nonlinear processes and numerous variables, with data-driven models becoming a key non-engineering approach to flood prevention and mitigation. Yet, single machine learning models are insufficient to fully capture the complex dynamics of the flood process. Study focus: We propose an interpretable flood forecasting hybrid model based on Transformer, LSTM, and Adaptive Random Search Algorithm (AGRS), termed as AGRS-LSTM-Transformer. Investigating the predictive performance of the hybrid model, this study compares it against AGRS-LSTM, AGRS-Transformer, AGRS-BP, and AGRS-MLP models, utilizing flood data from 1971 to 2013 years in the Jingle watershed. New hydrological insights for the region: The AGRS-LSTM-Transformer model demonstrates superior performance over benchmark models, achieving accurate runoff forecasts with a lead time of 1–6 hours. It achieves a Nash-Sutcliffe Efficiency (NSE) greater than 0.905, and root mean squared error (RMSE), mean absolute error (MAE), Bias, and relative error (RE) values for the runoff process below 34.891 m³/s, 25.125 m³/s, 9.537 %, and 8.025 %, respectively. The coupling between the LSTM layers and the Transformer input components plays a crucial role in the architecture of the AGRS-LSTM-Transformer model. Rainfall from stations situated near the main river channel and downstream flow sections exerted a positive influence. Runoff from the preceding moment significantly impacts the predicted flow, with the contribution of runoff inputs exceeding that of rainfall. Inputs nearer to the forecast moment do not invariably improve forecasting accuracy, and historical rainfall and runoff volumes with extended lag times may detrimentally impact the model prediction. The study highlights the potential of hybrid data-driven models in enhancing the accuracy of flood forecasting, offering insights for reducing uncertainty in flood prediction and interpreting machine learning flood forecasting models. This provides a scientific basis for flood early warning systems and water resource management.

Keywords

• Flood forecasting
• Rainfall-runoff
• Machine learning
• Interpretability
• AGRS-LSTM-Transformer
• Adaptive random search algorithm