The accuracy of traffic flow prediction significantly impacts the operation of Intelligent Transportation Systems (ITS). In this paper, we propose a Differential Time-variant (DT) Traffic Flow Prediction method, which can remarkably improve the accuracy and reduce the variance of traffic flow forecast based on deep learning models. To extract the temporal trend of the traffic flow at different locations, we apply data difference to preprocess the raw traffic data. This method can better eliminate the uncertainties of traffic flow series like volatility and anomaly. Then, time information is introduced in the form of One-Hot Encoding to effectively model the temporal patterns of traffic flow. Necessary analysis is presented to demonstrate the rationality. Three popular deep neural networks are applied to test our method, and experimental results on PeMS data sets indicate that it can make more accurate prediction compared with the same model.