Key points are not available for this paper at this time.
Accurate runoff forecasting plays a critical role in sustainable flood risk management and climate-resilient water resources planning, particularly in plain river-network basins, where runoff processes are influenced by strong temporal variability and intensive human regulation. To address the limitations of existing data-driven models in representing short-term temporal variability and long-range dependencies, this study develops a dual-branch temporal convolutional network–transformer encoder (TCN–TE) forecasting framework for multi-station runoff prediction. The proposed model integrates a temporal convolutional network (TCN) with channel attention (CA) to extract local temporal patterns and adaptively reweight multivariate hydrological features, and a gated recurrent unit (GRU)-enhanced transformer encoder (TE) to improve long-range temporal dependency modeling. In addition, an autocorrelation-based analysis is conducted to quantitatively determine the effective memory length of the runoff system, providing a statistically grounded basis for input window selection. The model is evaluated using daily runoff and rainfall data from the Dongtiao River basin in eastern China, including seven runoff stations and two rainfall stations over the period 2012–2024. Forecasting results under multiple horizons (1, 3, 7, and 14 days) demonstrate that the proposed TCN-TE model consistently outperforms representative deep learning baselines in terms of R2, RMSE, and MAE, with particularly significant improvements for medium- and long-term forecasts. The results suggest that the proposed model provides a useful data-driven multivariate forecasting framework for runoff prediction using multi-station hydrological observations.
Meng et al. (Mon,) studied this question.