Abstract: Snowmelt runoff is a crucial component of water resources in the arid regions of Xinjiang.Accurate runoff prediction is fundamental for effective water resource management.Using MODIS snow cover data and observed flow data from the Hetian River Basin in Xinjiang from 2001 to 2023, with snow cover percentage, snowline elevation, and large-scale meteorological and climatic indices as predictive factors, the dominant predictive variables were extracted by Principal Component Analysis (PCA).Then, three methods, including multiple regression analysis, support vector machine (SVM), and random forest (RF), were adopted to establish data-driven models for snowmelt runoff at two sections in the Hetian River Basin.Subsequently, based on the Stacking fusion algorithm, the above-mentioned models were integrated to establish an integrated forecast model for snowmelt runoff.The results indicate that all three models perform well in medium and long-term snowmelt runoff prediction, with the random forest model exhibiting superior accuracy compared to the multiple regression and support vector regression models.Based on the Stacking fusion algorithm, the integrated model, after fusing the multiple regression model, support vector machine model, and random forest model, outperforms single models, and the prediction accuracy is improved.Specifically, the root mean square error (RMSE) decreases from 0.308 m³/s to 0.240 m³/s, the mean absolute error (MAE) reduces from 0.227 m³/s to 0.188 m³/s, and the correlation coefficient (R²) increases from 0.864 to 0.874.The research results can provide a reference for water resources allocation and regulation, as well as flood disaster prevention in Northwest China.