Research on Hydrological Prediction Based on Improved Successive Variational Mode Decomposition and Dual-Attention Temporal Convolutional Network

To support water resource management, flood control, disaster mitigation, and environmental protection in the middle reaches of the Jinsha River, precipitation and runoff data from five monitoring stations—Shigu, A’hai, Zhongjiang, Jin’anqiao, and Panzhihua—were jointly used as multivariate inputs to forecast runoff at Panzhihua Station. To obtain well-separated intrinsic mode functions (IMFs) at different frequencies while avoiding the use of future information during feature construction, SVMD was implemented in a rolling manner, with only the historical observations available up to each forecasting origin used for decomposition. The original runoff series was reconstructed using sliding windows with different historical input window lengths, and the optimal historical input window length was selected according to the forecasting performance. For enhanced adaptive optimization and reduced manual intervention, an Improved Exponential–Trigonometric Optimization (IETO) algorithm was employed to optimize SVMD parameters. Finally, a dual-attention temporal convolutional network (DATCN) was developed for runoff prediction. At a historical input window length of L = 5, the proposed IETO-SVMD-DATCN model achieved an NSE of 0.9826, compared with 0.9813 for DATCN and 0.9820 for SVMD-DATCN. The proposed model also achieved lower RMSE, MAE, and MAPE values than the compared models. These results indicate improved forecasting performance for the evaluated runoff data.