Traffic Origin–Destination Flow Prediction Considering Individual Travel Frequency: A Classification-Based Approach

Abstract

Accurate prediction of traffic origin–destination (OD) matrices plays an important role in traffic management and urban development. The existing studies on OD matrix prediction primarily focus on time series prediction techniques, whereas the influences of individual activities have long been overlooked. To address this issue, a traffic OD flow prediction method that considers individual travel frequencies is proposed. The travel frequencies of the vehicles are determined using license plate recognition data. Based on travel frequency, all vehicles are classified into several categories using the K-means method. Subsequently, historical OD matrices for different vehicle categories are input into multiple deep learning models. These deep learning models are trained separately to predict the traffic OD matrices with respect to different levels of travel frequency. By aggregating these OD matrices, a short-term prediction of the total traffic OD matrices can be obtained. The proposed method is validated using real license plate recognition data collected from a specific area of Liuzhou City, China. The results demonstrate that the proposed method outperforms existing methods that do not consider travel frequency, with a reduction of 16.8% in mean absolute errors, a decrease of 16.2% in root mean square errors, and a 27.6% increase in R-squared values on average.