Machine learning-based flood forecasting models considering runoff process vectorization

Accurate multi-step-ahead flood forecasting is beneficial for flood control and disaster reduction efforts, as well as optimizing water resource management. This study proposed a runoff process vectorization method (RPV) and combined it with three machine learning (ML) models to construct the RPV-ML flood forecasting series models. The research area comprises three typical river basins : Gushanchuan River, Jialu River, and Zuli River, in the middle and upper reaches of the Yellow River. Rainfall-runoff data from 43, 28, and 37 flood events were respectively used for model training and validation, with a ratio of 7 ∶ 3 for flood events. The research shows that : ① Under the same lead time conditions, RPV-ML has higher Nash efficiency coefficients(E_NS), lower root mean square errors(E_RMS), and lower relative errors(E_R) of peak in flood forecasting for the Gushanchuan River, Jialu River, and Zuli River basins. RPV-ML outperforms ML models in terms of predictive performance, especially in lead times of 4-6 hours. ② The accuracy of RPV-ML and ML models gradually decreases as the lead time increases, but RPV-ML exhibits a slower decline in accuracy and demonstrates better robustness. ③ The RPV-improved Temporal Convolutional Network (TCN) can better overcome forecasting errors, and the RPV-TCN model performs the best in terms of predictive performance among the three basins. The research findings can provide the scientific basis for flood control and disaster reduction efforts in river basins.