Abstract: Based on the ECWMF model forecast products（2 m temperature,10 m wind,precipitation,etc.） and the 2 m temperature historical observation data of 80 national meteorological stations in Anhui province,three machine learning algorithms,decision tree（DT）,random forest（RF） and light gradient boosting machine（LightGBM）,were utilized to correct ECMWF model forecast products of the daily maximum and minimum temperature with the lead time of 0-72 hours.The corrected temperature products were further compared with that corrected by model output statistics（MOS）method and the forecaster’s subjective forecast products.The results show that:（1）The mean absolute error（MAE）of ECMWF daily maximum temperature forecast is obviously higher than that of minimum temperature,and MAE in mountainous area of Dabies and southern Anhui are large.（2）RF has the best performance in predicting daily maximum temperature and LightGBM in minimum temperature.Compared with the ECMWF model,the prediction accuracy has increased by 18.16% and 5.19%respectively.（3）The machine learning model fusing the surrounding stations information can effectively reduce the temperature forecast errors ofmountain meteorological stations.（4）Compared with the subjective correction,machine learning methods can significantly improve the temperature prediction accuracy of the numerical model in the transition weather,such as high temperature and cold wave.