HydroFusion ‐ GRU : A Hydrological Feedback‐Guided Deep Learning Framework for Short‐Term Daily Streamflow Forecasting

Reliable short‐term streamflow forecasting is essential for flood‐risk assessment and water‐resources management. This study proposes HydroFusion‐GRU, a hydrologically guided deep learning framework for daily discharge forecasting in the 4105.2 km 2 modeled catchment of the Batman River Basin, Türkiye. Using daily hydro‐meteorological data from 1981 to 2023, the model combines discharge memory, meteorological forcing, temporal attention, log‐residual learning, hydrological feedback, low‐flow protection, and spring‐runoff descriptors. The data were split chronologically into training (1981–2017), validation (2018–2021), and independent test (2022–2023) periods, with forecasts evaluated at t + 1, t + 3, and t + 7 horizons. At t + 1, HydroFusion‐GRU achieved MAE = 11.698 m 3 /s, RMSE = 32.279 m 3 /s, R 2 = 0.901, and KGE = 0.950, outperforming the best conventional benchmark, HistGradientBoosting (RMSE = 35.004 m 3 /s; R 2 = 0.889). At t + 7, HydroFusion‐GRU achieved RMSE = 52.699 m 3 /s and R 2 = 0.795, compared with 75.080 m 3 /s and 0.752 for HistGradientBoosting. Ablation analysis showed the greatest deterioration after wind removal and the smallest after precipitation removal.