Abstract: To reduce the spatiotemproral bias in precipitation forecasts in the Southwest region of China. This paper employs Quantile Mapping ( QM) and the machine learning method LightGBM to conduct error correction research on the May - August 0 — 45 day precipitation forecasts from the European Centre (EC) numerical models for the southwestern region . The results indicate that the high-value areas of root mean square error (RMSE) in the original EC model forecasts are mainly concentrated in eastern Sichuan to western Chongqing , and the spatial correlation descends rapidly with lead time . Although the QM method brings the probability density function curve closer to observations , it increases the forecast error of precipitation below moderate intensity due to the model's low spatial correlation coefficient , while the hit rate and the false alarm rate on rainy days have both increased . In contrast , the LGBM method significantly reduces RMSE and enhances spatial correlation coefficients across all forecast lead times , particularly demonstrating stronger generalization capability for moderate-to-heavy precipitation events . Further analysis shows that LGBM significantly outperforms both QM and the original forecasts in terms of reconstructing the probability density of precipitation , controlling RMSE of heavy precipitation , and identifying rainy- day events . This indicates that , compared to traditional statistical methods , the machine learning-based correction approach can effectively reduce spatio-temporal biases of precipitation forecasts for the southwestern region of China. This is a feasible path for future model correction research .