Abstract: This study proposes an improved stacked optimal screening machine learning model (REGOCF) to enhance the accuracy and stability of temperature prediction.The model integrates multiple factors within the study region,including terrain characteristics,underlying surface types,and forecaster experience,thereby reducing uncertainties arising from multi-model discrepancies.A comparative verification was conducted using nearly one year of forecast data from REGOCF,the China Meteorological Administration's Urban Guidance Forecast,ECMWF Thin,the National Intelligent Grid,and the Inner Mongolia Autonomous Region's RAP (Rapid Refresh) regional numerical model.The verification results demonstrate that the REGOCF algorithm exhibits superior performance in forecasting both maximum and minimum temperatures.Compared with single-model forecasts,the average accuracy improved by 5%-10%,with a marked reduction in forecast errors and a significant decrease in number of outliers,which is highly practical.Additionally,this study assessed the performance of the REGOCF algorithm across different seasons and climatic conditions,further corroborating its robustness and adaptability.This research not only provides an innovative approach to temperature correction but also provides a reference for the forecasting of other meteorological elements.