To address the challenge of decoupling spatiotemporal dynamics and topographic effects in climate modeling over complex terrain, this study proposes a ConvLSTM-XGBoost hybrid model based on dynamic Bayesian weighting. Using the Hongyuan Mountain region in Yunnan, China (22.5°-23.5°N, 102.5°-103.5°E) as a case study, the model achieves high-precision climate prediction. The ConvLSTM network captures the spatiotemporal evolution patterns (e.g., southwest monsoon front propagation) in the CN05.1 climate dataset at 0.25° resolution, while XGBoost quantifies the nonlinear modulation effects of 90-meter SRTM DEM topographic features (elevation, aspect) on precipitation phases. A Bayesian Model Averaging (BMA) framework is innovatively introduced to dynamically adjust model weights (XGBoost weight 0.68±0.05 during dry seasons, ConvLSTM weight 0.72±0.07 during monsoon periods), enhancing the model’s responsiveness to extreme events. Experiments based on climate data from 1961-2022 show that the hybrid model reduces the MAE in precipitation prediction by 30.5% compared to CMIP6 (0.0089), and improves the F1 score for identifying extreme precipitation events (>50 mm/day) by 20%. For temperature prediction, the model achieves a Tmax accuracy of 96.53% (error ≤3%), with a 52% reduction in high-value dispersion. This model provides a 1-kilometer resolution decision-making tool for mountain climate risk management, supporting drought warning and hydropower scheduling needs in Yunnan’s "Climate Adaptation Plan 2035", and offers a scalable framework for climate modeling in mountainous regions worldwide.