ABSTRACT Accurate runoff prediction is essential for flood control and water resource management yet remains constrained by the non‐stationarity and multi‐scale complexity of hydrological series. To address these challenges, this study proposes SFTformer, a hierarchical hybrid framework. Unlike conventional approaches, SFTformer utilizes a dual‐domain decomposition strategy—combining Seasonal‐Trend decomposition using Loess (STL) in the time domain with Frequency‐Adaptive Normalization (FAN) in the frequency domain—to effectively isolate high‐frequency stochastic residuals from stable seasonal patterns. To capture these decoupled features, we constructed a specialized TCN‐Informer architecture enhanced by the Bidirectional Gated Feature Fusion (BiGFF) mechanism. This module dynamically modulates the fusion of local temporal features and global dependencies, ensuring robustness across varying forecast horizons. Validated using daily runoff data from the Panzhihua II and Xiangjiaba stations in the lower Jinsha River basin, the framework demonstrates substantial improvements, particularly in medium‐term forecasting. Empirical results indicate that the model maintains an R 2 exceeding 0.94 even at the 14‐day horizon. In comparative benchmarks against the strongest baseline (Informer), SFTformer reduces RMSE, MAE, and MAPE by 26.5%, 24.1%, and 22.5%, respectively. These findings confirm that SFTformer effectively balances short‐term accuracy with medium‐term robustness, offering a reliable tool for real‐time flood risk management and sustainable water resource allocation in complex river basins.
Yan et al. (Tue,) studied this question.