Abstract: To address challenge of spatial temperature interpolation caused by complex terrain and uneven distribution of meteorological stations in Xinjiang Uygur Autonomous Region（abbreviated as Xinjiang）, a Stacking ensemble RF-SVR-MLP model based on Bayesian optimization has been proposed．Based on data from 84 meteorological stations in Xinjiang from 2011 to 2024, model has integrated multi-source topographic covariates such as elevation and slope．Random forest（RF）and support vector regression（SVR）have been employed as base models to explore nonlinear relationships, while multilayer perceptron（MLP）serves as meta-model to integrate prediction results．Bayesian optimization combined with K-5 fold cross-validation have been used to optimize hyperparameters．Multi-time scale validation showed that model's annual-scale mean absolute error was 1.105 °C, root mean square error was 1.337 °C, and coefficient of determination（R²）was 0.836．Compared to optimal single model, MAE and RMSE were reduced by 35.2% and 31.0%, respectively, while goodness of fit improved by 27.4%．In summer, model has resolved issues of single-model failure（R²=0.759）, and in winter, it has adapted to temperature inversion characteristics（reducing mean absolute error by 40.2%）．During critical period from April to October, mean absolute error consistently remained below 1.2 °C．This model has effectively broken through limitations of traditional methods and single machine learning models, accurately adapting to complex terrain and sparsely populated station areas．It has provided high-precision temperature data for agricultural planning and ecological protection in arid regions, thereby enriching methodological system for climate variables spatialization．