Highway Travel Time Prediction of Segments Based on ANPR Data considering Traffic Diversion

• Published in: Journal of advanced transportation Vol. 2021; pp. 1 - 16
• Main Authors: Du, Wenjun, Sun, Bo, Kuai, Jiating, Xie, Jiemin, Yu, Jie, Sun, Tuo
• Format: Journal Article
• Language: English
• Published: London Hindawi 2021; John Wiley & Sons, Inc; Hindawi Limited; Wiley
• Subjects: Accuracy; Artificial neural networks; Cable television broadcasting industry; Deep learning; Feature extraction; Forecasts and trends; Freeways; Highways; Information systems; Long short-term memory; Methods; Neural networks; Predictions; Roads & highways; Segments; Time series; Traffic congestion; Traffic flow; Traffic information; Traffic management; Traffic models; Transportation; Travel; Travel time; United States; China
• ISSN: 0197-6729; 2042-3195
• DOI: 10.1155/2021/9512501

Travel time is one of the most critical parameters in proactive traffic management and the deployment of advanced traveler information systems. This paper proposes a hybrid model named LSTM-CNN for predicting the travel time of highways by integrating the long short-term memory (LSTM) and the convolutional neural networks (CNNs) with the attention mechanism and the residual network. The highway is divided into multiple segments by considering the traffic diversion and the relative location of automatic number plate recognition (ANPR). There are four steps in this hybrid approach. First, the average travel time of each segment in each interval is calculated from ANPR and fed into LSTM in the form of a multidimensional array. Second, the attention mechanism is adopted to combine the hidden layer of LSTM with dynamic temporal weights. Third, the residual network is introduced to increase the network depth and overcome the vanishing gradient problem, which consists of three pairs of one-dimensional convolutional layers (Conv1D) and batch normalization (BatchNorm) with the rectified linear unit (ReLU) as the activation function. Finally, a series of Conv1D layers is connected to extract features further and reduce dimensionality. The proposed LSTM-CNN approach is tested on the three-month ANPR data of a real-world 39.25 km highway with four pairs of ANPR detectors of the uplink and downlink, Zhejiang, China. The experimental results indicate that LSTM-CNN learns spatial, temporal, and depth information better than the state-of-the-art traffic forecasting models, so LSTM-CNN can predict more accurate travel time. Moreover, LSTM-CNN outperforms the state-of-the-art methods in nonrecurrent prediction, multistep-ahead prediction, and long-term prediction. LSTM-CNN is a promising model with scalability and portability for highway traffic prediction and can be further extended to improve the performance of the advanced traffic management system (ATMS) and advanced traffic information system (ATIS).