Urban Short-Term Traffic Flow Prediction Based on Stacked Autoencoder

Traffic prediction technology is important for intelligent transportation systems. To improve the performance of traffic prediction, this paper proposes a deep learning method for short term traffic prediction. The ground traffic data from microwave sensors at 2nd Ring road, Beijing, China, was selected as the candidate dataset. A stacked auto-encoder neural network (SAE-DNN) model is introduced to forecast short term traffic conditions. First, the SAE model is applied to extract inherent information within historical raw data. Second, output of the SAE model is used as the input of DNN model to perform the calibration and validation of DNN model. Finally, the optimal prediction result can act as a referenced traffic condition in the next period. Testing results show that the SAE-DNN model is an obvious improvement of traffic prediction compared to the traditional back-propagation neural network (BPNN).