Travel Time Prediction by Providing Constraints on a Convolutional Neural Network

Abstract

Traffic prediction is a complex, nonlinear spatiotemporal modeling problem. Data-based methods have been widely applied to traffic information prediction because of their ability to model the complex, nonlinear spatiotemporal traffic dynamics. These abilities are based on statistical methods and machine learning methods to extract features from historical traffic information. Despite the success of these methods, the local-level representation in the data-based methods is clearly limited because it does not directly reflect the relationship between each historical data point in the input sequence and the predicted information. Convolutional neural networks (CNNs) have also been widely applied in many other fields because of their feature extraction ability, which greatly enhances the learning ability of the standard, fully connected layers of neural networks by providing constraints in the task domain. Thus, we create a global-level representation to reflect the relationship between each historical traffic data point and the predicted information by using the proposed window that provides constraints in the task domain and exploits a max-pooled CNN to automatically extract global-level features for traffic information. To the best of our knowledge, our proposed method is the first example of using the new window to provide constraints in the task domain for CNNs. The proposed method is trained in an end-to-end way by the back-propagation approach in which the AdaGrad method is used to update the parameters of the proposed method. Based on the data set provided by Highways England, the experimental results show that the proposed method outperforms all baseline methods in terms of accuracy.