A multi-model integrated forecast correction process and scheme are constructed for 2 m temperature forecasts at Xi’an Station based on the 2021-2024 European Centre for Medium-Range Weather Forecasts (ECMWF) model forecasts and the 2 m temperature observations from Xi’an Station. The data from 1 September 2021 to 31 December 2023 are used as the training set for factor screening, parameter tuning, and model ensemble, while the data from 1 January to 30 April 2024 are taken as the test set to assess the forecast performance of numerical models and models trained under different schemes. Through subjective experience screening and time-lag correlation analysis, seven model forecasting physical variables closely related to temperature changes, as well as different lead time high level key zone variables and other characteristic factors get optimized. XGBoost, LightGBM and CatBoost are used for single model bias correction, and finally model fusion optimization is achieved through Stacking ensemble. The results show that ECMWF model exhibits a systematic cold bias in temperature forecasts at Xi’an Station, with the error being significantly greater at night than during the day and the cold bias intensifying during cooling and precipitation processes. After Bayesian optimization and parameter tuning, all the three machine learning models are able to effectively correct mode bias with root mean square errors (RMSE) reduced by 0.039℃, 0.030℃, and 0.027℃, respectively. Subsequent feature factor optimization further improves the single model forecast accuracy by approximately 0.257℃. The Stacking ensemble surpasses the traditional weighted ensemble. After ensemble, the RMSE of temperature forecasts is reduced by 0.023℃, and the forecast accuracy within 2℃ is improved by 2.589%. During the significant cooling and precipitation process, the forecast RMSE has a maximum reduction of 0.481℃ compared to that by the single model.