Using the idea of climate probability statistics and multi-time average, the prediction bias of 2 m temperature products of ECWMF high-resolution model in Liupanshui from 2018 to 2019 were analyzed, and the prediction accuracy of the model revised by the index in 2020 was tested. The results show that the temperature prediction bias of ECWMF high-resolution model in Liupanshui area decreases gradually with the increase of aging, and the average annual prediction bias and standard deviation of the maximum temperature of each aging are significantly higher than the minimum; the model has the highest reliability in early summer (June) and lowest in spring (March-April) in Liupanshui; by adopting the maximum proportion of prediction bias to the model minimum temperature forecast which monthly multi-time averaged for correction, and adopting the different correction method and correction index to correct the model 24 h maximum temperature forecast according to the weather type, the prediction accuracy of the comprehensive minimum temperature in the next five days (120 h) and the maximum temperature in 24 h in Liupanshui area can be greatly improved which near are 90% and above 70% respectively; after revision, the prediction accuracy of the average minimum temperature in 2020 in Liupanshui area is equivalent to the actual situation, while the prediction accuracy of the maximum temperature in 24 h is higher than the actual prediction accuracy.