Aiming at the problem that agricultural greenhouse temperature measurement is greatly affected by noise and is not suitable for direct prediction, an integrated prediction model XGB-KF, which combines XGBoost and Kalman Filter, was proposed. The model firstly estimated the temperature of the greenhouse at the current moment based on XGBoost, and then Kalman Filter was used to dynamically modify the estimated result. Finally, the prediction result is obtained. Based on the sensor data of agricultural greenhouse in Zhuozhou, the numerical experiment was carried out. At the same time, root mean square error (RMSE) was used as the main indicator to evaluate the model. Compared with XGBoost, Bi-LSTM and Bi-LSTM-KF methods, the root mean square error (RMSE) of XGB-KF model was reduced by 5.22%, 10.85% and 7.45% respectively.