Proper prediction of traffic congestion is an essential component of many intelligent transportation systems (ITS). Multi-layer feed-forward (MLF) networks were commonly used for this purpose. However, MLF models fail to fully capture the temporal relationship among traffic series. In this paper, we present a traffic congestion prediction method based on partially recurrent neural networks (RNNs). In recurrent neural networks, the temporal relationship of the traffic series data is explicitly modeled via internal states. By using sensors embedded in the expressway, Elman RNN models have been developed to predict future volume and occupancy values 1, 5, 10 and 15 minutes in advance given current sensor data. A comparison between recurrent and traditional MLF neural networks was performed and the results suggest that RNN outperformed MLF with up to 5 percent improvement of prediction accuracy.
Abstract