Aiming to predict apple production in Loess Plateau in a timely and accurate manner, based on historical weather records in a total of 86 counties in the apple producing areas of the Loess Plateau, the data of different meteorological feature variables (e.g., temperature, precipitation, and radiation in different apple growing months), spatial feature variables (e.g., latitude, longitude and elevation of meteorological stations), and meteorological disaster feature variables (e.g., time of freezing damage at flowering stage, time of continuous rain, and standardized precipitation evapotranspiration index (SPEI)) were extracted at first. The influential feature factors were determined according to Spearman correlation analysis. Nextly, the prediction models for apple relative meteorological yield were established based on different algorithms (i.e., gradient boosting decision tree, GBDT;support vector machine, SVM;Bayesian regularization back propagation artificial neural network, BRBP;multiple linear regression, MLR). At the same time, the optimal combination of model input feature variables was determined for each of the established yield prediction models. Finally, based on the optimal combinations of input feature variables in different apple growth periods and in different months of apple growing seasons, the prediction leading time were analyzed for different simulation models for apple relative meteorological yield. The results were as follows: the influential meteorological feature variables were the highest temperature, lowest temperature, air relative humidity, precipitation and solar radiation. The best model input variable combination was selected as the influential meteorological, spatial and disaster feature variables. Based on the best combination of model input variables, the GBDT and BRBP models had better prediction accuracy (r was 0.77, RMSE was 0.44;r was 0.70, RMSE was 0.44), while the MLR model performed the worst (r was 0.63, RMSE was 0.49). In different growth periods of apples, the GBDT and BRBP models could obtain relatively high apple yield prediction accuracy in each growth period, while the SVM and MLR models could obtain relatively ideal simulation results in apple fruit expansion period. In each month of the apple growing season, the GBDT, SVM, BRBP and MLR models could realize early prediction of apple relative meteorological yield about one to two months before apple maturity. The research result can provide a scientific foundation and technical reference for apple yield prediction on the Loess Plateau.