Purple soil is one of the unique soils in China. Because of its low degree of soil development and poor soil erosion resistance, soil erosion in the purple soil area is very serious, which seriously hinders local agriculture and economic development. 【Objective】 To accurately predict soil erosion changes, we need to scientifically and rationally prevent the occurrence of soil erosion, provide a reference for the prevention and control of soil erosion on purple soil slopes, and promote research on soil and water conservation in purple soil areas. 【Method】 Based on the rainfall data of 43 typical purple soil run-off communities, rainfall, duration and I30 are used as characteristic indicators, and the K-means clustering algorithm was used to classify rainfall; the Random forest (RF) algorithm was used to evaluate the importance of each influencing factor on the depth of run-off (H) and soil loss (S), and the important factor screening was carried out in combination with the path analysis method. The key factors obtained from the screening were used as the input variables of the model, with H and S as the outputs. Also, a predictive model was constructed using BP neural network. 【Result】 The results showed that:(1) The erosive rainfall in the purple mound area in the southwest can be divided into three categories. Among them, the A rain type had a long duration, medium rain intensity, heavy rain volume, the lowest frequency of occurrence, and the greatest erosion, which was an important rain type that caused slope erosion in the area; the B rain type had a short duration, heavy rain, and light rain volume, and the most frequent occurrence, which was the main rain type that caused slope erosion in the area; and the C rain type, which had a medium duration, medium rain intensity, and medium rainfall, which occurred more frequently and also made a greater contribution to slope erosion in the area. (2) The importance of characteristic factors to H and S under each rain type was significantly different. For rain type A, the rain duration (T) was the most important to S (31%), and the importance of each factor to H was relatively uniform. Under rain type B, the influence of each factor on S was small, and the rainfall erosion force (F) was the most important to H (29%). Also, the degree of importance was the highest (29%). Under the rain type C, rainfall (Pr) was the most important to H and S(33%, 36%). (3) The H and S under the three rain types were significantly affected by Pr and the H and S under the B and C rain types were most significantly affected by F, vegetation cover (Vs), and average plant height (Ph) at the same time. (4) The prediction accuracy of H and S using the BP neural network was high, the efficiency coefficient of Nash-Suttclife is higher than 0.95, and the accuracy of the prediction model for H was higher than that of the prediction model for S. 【Conclusion】 Research on the purple soil mound area in the southwest needs to focus on preventing soil erosion caused by high-frequency heavy rainfall. In this study, the two models that used rainfall and other factors to calculate H and S had high forecasting capabilities, which provide technical support for soil erosion prediction to achieve accurate forecasting of soil erosion.