Stretch-wide traffic state prediction using discriminatively pre-trained deep neural networks

The paper adopts the state-of-the-art machine learning deep neural network to model the evolution of the traffic state along a 21.1 miles long stretch of the I-15 highway. The built model is used for short-term prediction of the traffic states. The 21.1 miles stretch is divided into 43 segment. Building a predictive model for this stretch is a multivariate problem where the responses are the speeds/flows for different road segments at different time horizons. Considering traffic state short-term prediction as a multivariate problem ensures that all the spatiotemporal correlations are maintained. We adopt the deep neural network to predict the traffic state stretch-wide for up to 120 minutes in the future. Due to the required large computation time and memory to train the deep neural network we used the divide and conquer approach to divide the large prediction problem into a set of smaller overlapping problems. These smaller problems can be solved using a medium configuration PC in a reasonable time which makes the proposed technique suitable for practical applications. Furthermore, the prediction results showing a superior prediction when compared to prediction results obtained using the partial least squares regression.