A CNN-LSTM-Attention fusion model for medium to long range hydrological forecast and its application in the Dongjiang River Basin
PENG Haibo
DAI Shanjin
LI Zehua
ZHONG Hua
LI Yi
LIU Jing
TIAN Zhaowei
XU Fei
Abstract：To address the challenges of long lead times,complex influencing factors,and high uncertainty in medium-to long range hydrological forecast,this study proposes a deep learning-based hydrological forecast method.The method extracts time-varying features of antecedent climate indices using convolutional neural networks(CNN),captures temporal dependencies of hydrological processes with long short-term memory networks(LSTM),and focuses on key abrupt signals through a multi-head attention mechanism(Multi-Head Attention),thereby dynamically correcting the European Centre for Medium-Range Weather Forecasts(ECMWF)SEAS5 numerical forecast products.Taking the Dongjiang River Basin as a case study,a CNN-LSTM-Attention deep learning fusion model was constructed using monthly hydrological and meteorological observation data from 1993 to 2022.The results show that for lead times of 1～3 months,the model significantly improves precipitation forecast accuracy,reducing the mean absolute error(MAE)by 0.6～5.5 mm and increasing the coefficient of determination(R²)by 0.10～0.11 compared to the SEAS5 ensemble mean forecast.In terms of runoff,the MAE is reduced by 0.4～5.7 m³/s and R² is improved by 0.07～0.31 for lead times of 1～5 months.Additionally,the developed model integration and visualization platform offers an operationally valuable solution for medium-to long range hydrological forecast in the Dongjiang River Basin.
Keywords：Dongjiang River Basindeep learningconvolution neural networklong short-term memoryattention mechanism
Publication Date：2025-04-30
Online Publishing Date：2025-08-15(First online date of this platform, not the publication date of the document)
Pages：7( 61-66,72 )