Global warming has made water resources more uneven in space and time, making water management harder. This study used the HBV-XGBoost model to see how climate change affects runoff in the Xiangxi River Basin. The HBV model simulated water processes, and XGBoost improved predictions by handling complex relationships. This study used the SDSM to create climate data for 2025–2100 and looked at runoff trends under different emission scenarios. The HBV-XGBoost model performed better than the HBV model in simulating runoff. Future predictions showed big differences in runoff trends under various SSP scenarios in the 2040s and 2080s. For example, under SSP585, the ACCESS-CM2 model projected a May runoff increase from 1527.52 m³/s to 2344.42 m³/s by the 2080s, and ACCESS-ESM1-5 projected an increase from 1462.11 m³/s to 2889.58 m³/s. All GCMs predicted a large rise in annual runoff under SSP585 by the 2080s, with FGOALS-g3 showing the highest growth rate of 76.54%. The model accurately simulated runoff changes and provided useful insights for adapting water management to climate change. However, this study has limitations, including uncertainties in machine learning models, potential input data biases, and varying applicability under different conditions. Future work should explore more climate models and downscaling methods to improve accuracy and consider local policies to better address climate impacts on water resources.

Accurate runoff simulation and prediction are crucial for water resources management, especially under the impact of climate change. In this study, a multi-physics ensemble Bayesian model averaging (MPE-BMA) model is developed to improve runoff prediction accuracy by integrating a soil and water assessment tool (SWAT), hydrologiska byråns vattenbalansavdelning (HBV) model, and Bayesian model averaging (BMA) into a general framework. The MPE-BMA model integrates the strengths of the SWAT and HBV models. This approach enhances the robustness of simulation outputs and reduces uncertainties from single-model methods. MPE-BMA is subsequently employed to simulate and predict runoff for the upper reaches of Xiangxi River Basin (XXRB) in China, where four general circulation models (GCMs) and three shared socioeconomic pathways (SSP126, SSP245, and SSP585) are considered. Multiple statistical metrics (R², NSE, and RMSE) prove that the MPE-BMA model outperforms the single models of SWAT and HBV. Results reveal that higher-emission scenarios generally lead to significant decreases in runoff, particularly by the 2080s. Specifically, under SSP585, runoff is projected to decrease by approximately 4.61–12.68% by the 2040s and 5.96–11.28% by the 2080s compared to the historical period. From the perspective of monthly and seasonal runoff changes, the peak runoff is projected to shift from June to May by the 2080s. Additionally, under SSP585, spring and summer runoffs tend to significantly increase, while winter runoff decreases sharply, leading to wetter summers and drier winters. These findings underscore the importance of enhancing water use efficiency, upgrading hydropower stations, and implementing watershed management practices to ensure sustainable water resources management in the XXRB amidst climate change.