In the current research of traffic flow prediction, the time step mainly depends on the artificial selection which is easy to be disturbed, and there is a lack of method to select the time step from the theoretical aspect. In order to adaptively select the time step, three typical algorithms of least squares support vector machine (LSSVM), random forest (RF) and long short-term memory (LSTM) in machine learning are used to predict traffic flow with multiple time step based on the autocorrelation analysis of traffic flow historical time series and explore the feasibility of selecting the best time step by the value of autocorrelation coefficient. The experimental results show that when the autocorrelation coefficient of the time step is 0.83-0.91, LSSVM can obtain better prediction accuracy while the autocorrelation coefficient is 0.47-0.51, LSTM can have better prediction accuracy. However, due to the low autocorrelation degree of the time step corresponding to the lowest error of traffic flow prediction, the autocorrelation analysis method may not be applicable to RF.