This paper constructed a model for short term forecasting of station temperature based on causal analysis of information flow and machine learning algorithms， taking Songshan Station in Taiwan as the test samples， using ECMWF reanalysis data， measured meteorological data of Songshan Station and CLDAS-v2.0 near-real-time product data， combined with causal analysis and correlation analysis， using four Machine Learning (ML) algorithms: BP neural network， Random Forest (RF)， Least Squares Support Vector Machine (LSSVM) and Bayesian Network (BN) to carry out temperature short-term forecasting experiments.The conclusions are as follows: (1)driven by any of the data selected in this paper， for BP neural network， RF， and BN， the prediction results based on causal analysis are better than those of correlation analysis， and the average reduction of RMSE is between 1%-2%. For LSSVM， the difference between causal analysis and correlation analysis is small， which verifies that causal analysis has better correlation mining capabilities; (2) adding adjacent spatial prediction factors in the forecast model can improve the temperature prediction effect， and the average reduction of the RMSE of the improved temperature prediction model is between 2%-8%; (3) in the case of small samples， the forecasting effect based on CLDAS-V2.0 data is better than that of ECMWF reanalysis data， and the average reduction of RMSE is between 4%-8%， which verifies that the quality of CLDAS-V2.0 data in China is indeed better than that of similar international products.