Abstract : To address the challenge of forecasting water inflow in small watersheds lacking historical observation data and to overcome the limitations of traditional hydrological models that require long-term time series and purely data-driven models constrained by data quality and feature selection , this article proposes an hourly-scale inflow forecasting method based on a physically-informed Bidirectional Gated Recurrent Unit ( Bi-GRU) deep learning framework . The Liuxihe Reservoir watershed is taken as a study area. The key topographic features is firstly extracted from a high-resolution ( 12 m ) digital elevation model ( DEM ) . Then , the issue of uneven spatial distribution of meteorological data inputs is addressed by a combination of grid-based interpolation and the Random Forest algorithm. A physical constraint mechanism is integrated into the model to guide feature selection , ensuring hydrological consistency. Finally , the Bi-GRU network is trained to capture complex temporal dependencies during the runoff generation . The experiment results show that the proposed approach can achieve high predictive accuracy , with Nash-Sutcliffe Efficiency (NSE) coefficients reaching 0 . 90 during low-flow periods and 0 . 96 during high-flow periods . This study demonstrates a promising solution for inflow prediction in data-scarce basins and has an important application value in the operation of hydrological forecasting.