To address the limitations of data-driven model, which lack physical constraints, and to enable rapid prediction of reservoir thermal structure under future non-stationary climate scenarios, this study proposes a hybrid framework that integrates physical mechanisms with data-driven modeling for predicting reservoir water temperature profiles. Utilizing observed data from the Sanbanxi Reservoir (2007-2016), a one-dimensional hydrodynamic-water temperature model (GLM), a Random Forest-Bidirectional Long Short-Term Memory (RF-BILSTM) machine learning model, and Global Climate Models (GCMs), we reconstructed the historical thermal structure of the reservoir and projected its evolution under future climate scenarios from 2023 to 2100. The results indicate that: (1) A physically constrained training data set generated by combining the GLM model with measured data effectively mitigates spurious correlations inherent in purely data-driven approaches. The RF-BILSTM prediction framework achieved high simulation accuracy, with R2> 0.9. (2) Projections from GCMs show a significant increasing trend in future air temperature at the Sanbanxi Reservoir (p < 0.01). The temperature increase under the SSP5-8.5 scenario is substantially greater than under SSP2-4.5, while precipitation also exhibits a significant upward trend. (3) By 2100, the average water temperature in the reservoir is projected to rise by 0.39 ℃ and 0.87 ℃ under the SSP2-4.5 and SSP5-8.5 scenarios, respectively, showing pronounced vertical differentiation. Surface warming is more significant, with projected increases of 0.94℃ and 2.04℃, respectively. The annual mean surface-bottom temperature difference is expected to increase by 1.32℃ and 1.55℃, with the maximum vertical difference occurring in August. Annual water column stability (Schmidt stability, St) is projected to increase by 643.41 J/m2(+7.85%) and 1829.47 J/m2(+22.31%), respectively, indicating intensified thermal stratification. This could elevate the risks of surface algal blooms and nutrient stratification enrichment, while also advancing the spawning timing of downstream fish. This study presents a novel approach that deeply integrates physical mechanisms with data-driven techniques, providing technical support for reservoir water temperature management.