Abstract: Based on the European Centre for meaium-Range Weather forecasts(ECMWF) model with a horizontal resolution of 0.1°×0.1° and the observation data from 20 national automatic weather sites in Beijing from July 1,2019 to March 15,2024,the characteristics of the model for 2 m temperature bias forecasted by ECMWF model during the historical heating season were analyzed.Extreme random forest,decision tree,gradient boosting tree,linear regression,and Lasso regression methods were used to correct the 2 m temperature forecasts from ECMWF model.The results show that the overall 2 m temperature forecast during the historical heating season in 2019-2023 in the urban area of Beijing by ECMWF is low,with the largest bias occurring in the afternoon and with the average bias of -2.3 ℃.While the 2 m temperature forecast in the suburban area is low in the morning and high in the afternoon,with the largest positive and negative bias occurring at 07:00 and 16:00,being 1.7 ℃ and -2.2 ℃,respectively.After the correction by machine learning method,the mean bias and root mean square error of urban and suburban sites in Beijing in the 2023 heating season (from November 7,2023 to March 15,2024) are significantly decreased,in which the correction effect of the extreme random forest is the best,and the improvement rates of root mean square error in the urban and suburban areas are 24.2% and 35.4%.After the correction by machine learning method,the accuracy of daily mean temperature forecast bias within ±0.5 ℃,±1.0 ℃ and ±2.0 ℃ at 9 sites in urban area of Beijing in the 2023 heating season is significantly improved,with the maximum improvement rates of 31%,44% and 40%,respectively,and the extreme random forest and decision tree have the best performance.