Nowadays, most temperature prediction models are based on numerical prediction. However, there is a major problem with this model, that is, the prediction accuracy is completely affected by the accuracy of numerical prediction, resulting in excessive reliance on the model and lack of knowledge of the actual weather data. In order to solve this problem, based on the data of automatic weather stations in Guizhou Province from 2013 to 2022, the Pearson correlation coefficient between the daily maximum and minimum temperatures in Xixiu District of Anshun City and other stations in the province was obtained by using the summer temperature data on the basis of considering the correlation between the upstream and downstream temperatures. Then, machine learning blocks were used to select the Stacking regression model and establish a local future 24 h temperature forecast method. The results show that: (1) the correlation between maximum and minimum temperature in the upper and lower reaches of Xixiu District passes the significance test of 0.005, which indicates that the 24-hour temperature change in Xixiu District is mainly affected by Bijie, Dafang, Bozhou, Kaiyang and Guiyang in the upper reaches; (2) The regression model established in this paper can well predict the trends of maximum and minimum air temperature in 24 h. Under the ±2 ℃ temperature inspection method, the accuracy rate is 83.7% and 93.47%, respectively. (3) The prediction accuracy of maximum temperature is lower than that of minimum temperature, which also reflects the difficulty of maximum temperature forecast in Xixiu District. This method can effectively reduce the over-reliance on numerical models, and has high accuracy, stability and universality in predicting local 24 h temperature.