Water erosion causes soil fertility loss and land degradation, posing serious threats to agricultural production, soil and water conservation, and environmental protection. Accurately assessing the spatiotemporal dynamics and driving mechanisms of water erosion is essential for mitigating regional erosion risks. In this study, the Revised Universal Soil Loss Equation (RUSLE) was coupled with the Transport Limited Sediment Delivery (TLSD) to simulate water erosion in the Shiyang River Basin from 2001 to 2020. Additionally, the Random Forest (RF) algorithm was employed to quantify the contributions of different driving factors to net soil erosion. The results showed that the RUSLE-TLSD model reliably simulated soil erosion processes in the basin (NSE = 0.70). The annual mean net soil erosion rate varied between 2.28 and 9.68 tha-1a-1, with an overall decline, and the most intense water erosion occurred during the summer (June-August). Areas of intense water erosion were primarily concentrated in the southern Qilian Mountains, characterized by steep relief, abundant rainfall, and strong sediment transport capacity. The RF model explained approximately 85% of the variance, indicating that LUCC, NDVI and slope had the most significant impact on the spatial pattern of water erosion in the basin, and the area with sparse vegetation and large topographic relief had a high risk of water erosion. It is hoped that the findings of this study will provide a reference for water erosion risk assessment and management planning in arid zone basins.