Evaluation of climate prediction models in Yunnan, China: traditional methods and AI approaches

Understanding regional climate variability is essential for effective climate risk management, particularly in areas with complex terrain like Yunnan Province, China. Traditional regional climate models (RCMs), such as RegCM, face limitations in predictive accuracy and computational efficiency due to their reliance on nonlinear physical simulations. To address these challenges, this study introduces a comprehensive framework to evaluate regional climate predictions using artificial intelligence (AI) models. Specifically, we assess the performance of five mainstream AI models-CNN, LSTM, Transformer, CNN-LSTM, and LSTM-Transformer-in predicting key climate variables: temperature, precipitation, and relative humidity. Daily meteorological observations from 25 national stations (2004-2018) were employed, with dimensionality reduction and temporal feature encoding enhancing the sequence-based learning models. Model performance was evaluated using RMSE, MAE, and Pearson correlation coefficient (R). The results demonstrate that AI models substantially outperform RegCM, particularly for temperature and humidity predictions. Among them, the LSTM-Transformer achieved the highest accuracy in temperature (RMSE = 0.7410, R = 0.9938) and humidity (RMSE = 3.7054, R = 0.9710), while CNN-LSTM was most effective for precipitation (RMSE = 4.7260, R = 0.8559). These findings highlight the potential of artificial intelligence for advancing multivariate climate prediction in regions with significant spatial heterogeneity, providing a data-driven basis for more accurate climate risk assessment and early warning applications.