Abstract: This study addresses the challenges posed by complex hydrological conditions and limited accuracy in runoff simulations in the source region of the Yellow River. We develop a runoff simulation method that integrates potential evapotranspiration （PET） predictions to improve the reliability of runoff modeling in alpine areas. Three machine learning approaches—Random Forest （RF）, Multilayer Perceptron （MLP）, and Extreme Learning Machine （ELM）—were employed, along with deep learning models LSTM and PatchTST. These models incorporated PET predictions for runoff simulations and evaluated PET simulation performance under different combinations of meteorological factors. The key findings are as follows: ① Maximum temperature emerged as the primary driver of PET simulation, with optimal accuracy achieved by combining maximum temperature, relative humidity, and wind speed. ②Among the deep learning models, although PatchTST underperformed compared to LSTM in predicting one-month-ahead PET, it exhibited superior performance in multi-step forecasting. ③Model performance improved significantly after the integration of PET predictions. For example, the Nash-Sutcliffe efficiency coefficient for the PatchTST model at Tangnaihai station increased from 0.706 to 0.896 （a 26.9% improvement）. The mean absolute percentage error decreased from 23.502 to 18.305 （a 22.1% reduction）, while the root mean square error dropped from 276.7 to 160.8 （a 41.9% reduction）. These results indicate that PET data effectively capture the dynamic impact of evapotranspiration on runoff loss, providing a more precise solution for hydrological forecasting in data-scarce alpine regions.