To explore the impact of different prediction schemes on the runoff prediction of machine learning models, the interval watershed (Wangjiaba~Jiangjiaji~Runheji) of Huaihe River Basin was taken as an example. In this study, we designed seven runoff prediction schemes and used three machine learning models —— LSTM (Long Short-Term Memory neural network), RF (Random Forest) and SVR (Support Vector Regression) to predict the runoff in the interval watershed. The results show that: (1) the sensitives of three machine learning models to rainfall information is different. The best scheme is that considering both runoff influencing factors and prior historical runoff. However, the importance of prior historical runoff diminishes when extending lead time. (2) The performance of the three machine learning models for runoff prediction is different with different lead time. The three machine learning models all perform well with 1-day lead time, SVR performs better with 2~4 days lead time and better performance for RF when lead time is 5~7 days. The study can provide a reference for the runoff prediction based on the machine learning.