A multi-scale deep learning framework for medium-long-term streamflow forecasting based on EMD-TCN-GRU

Accurate medium-long-term streamflow forecasting underpins flood mitigation and water-resource management across the Yangtze River Basin. Single deep-learning approaches remain challenged by non-stationarity, intricate long-range dependencies, and extreme-event sparsity. We propose an EMD-TCN-GRU framework: empirical mode decomposition (EMD) decomposes daily discharge into quasi-stationary modes and a residual, temporal convolutional networks (TCN) extract multi-scale temporal features in parallel, and gated recurrent units (GRU) generate multi-step forecasts for each component prior to linear recombination. When the forecast horizon is set at 3 days, the model—trained on 2013-2022 Wuhan observations—records an R2 of 0.9951 and a MAPE of 2.87%. Extending to 7 days, the R2 is 0.9925 and the MAPE is 3.22%. At 15 days, the R2 remains at 0.9922 while the MAPE is 3.25%. Compared to a standard GRU, the MAE is reduced by 63%, 51%, and 58%, respectively, and performance decay over time is negligible. Systematic ablation studies corroborate that the decomposition-convolution-gating pipeline is the primary factor in the observed increase in accuracy. The elimination of EMD serves to amplify residual noise, while the removal of TCN results in the severing of long-range information pathways. Furthermore, the substitution of multi-step GRU forecasting with single-step GRU forecasting triggers rapid error accumulation. The framework provides a robust, transferable solution for real-time flood warning and medium-long-term water allocation in the Yangtze River and analogous complex networks.