Revolutionizing flood forecasting by integrating rainfall-runoff correlation analysis with advanced deep learning techniques

Dong-mei Xu, Zhan Xu, Wen-chuan Wang, Yan-wei Zhao, Hong-fei Zang,
Revolutionizing flood forecasting by integrating rainfall-runoff correlation analysis with advanced deep learning techniques,
Results in Engineering,
Volume 28,
2025,
107804,
ISSN 2590-1230,
https://doi.org/10.1016/j.rineng.2025.107804.
(https://www.sciencedirect.com/science/article/pii/S2590123025038563)
Abstract: The paper aims to enhance the accuracy and reliability of flood forecasting by integrating advanced deep learning techniques with hydrological principles. Specifically, it proposes a novel LSTM-Transformer flood forecasting model (RRCALT) based on rainfall-runoff correlation analysis (RRCA) to improve the determination of input time steps and overall prediction performance. The study develops the RRCALT model, which combines Long Short-Term Memory (LSTM) networks and Transformer architecture. The model uses RRCA to determine the optimal input time step for the deep learning model, ensuring it captures the complete evolution of the rainfall-runoff process. The model's performance is validated using data from the Luohe River basin and the Dangze River basin, comparing it with LSTM, Transformer, and distributed physical models. The main contribution is: (1) The RRCALT model integrates LSTM and Transformer through RRCA, enhancing the understanding and prediction of complex flood time series. (2) A novel strategy for determining the input time step of deep learning models in flood forecasting based on RRCA is introduced, providing a more scientific and rational determination of the time step. (3) The model demonstrates strong adaptability and practical value in different hydrological environments, advancing flood forecasting technologies. The RRCALT model significantly outperforms other models in terms of accuracy and robustness. Compared to the Transformer model, it reduced MAE by 66.7% and 67.9%, RMSE by 71.9% and 71.1%, and increased KGE by 6.62% and 12.41% in the two study areas. The model also performed well in event-based flood evaluation standards, showing superior performance in predicting flood peaks and volumes. The study provides a robust and innovative approach to flood forecasting by combining the strengths of deep learning and hydrological principles. The proposed method offers a more accurate and reliable solution for flood prediction.
Keywords: Rainfall-runoff; Flood forecasting; LSTM; Transformer; Correlation analysis; Input time step