Short-term prediction of traffic state: Statistical approach versus machine learning approach

Abstract

Short-term traffic prediction helps intelligent transportation systems to manage future travel demands. The objective of this paper is to predict the state of traffic in the case of Karaj to Chaloos, a suburban road in Iran. To this end, two approaches, i.e., statistical and machine learning, are employed. In addition, the performance of the multinomial logit model is evaluated using Support Vector Machine (SVM) and Deep Neural Network (DNN) as two top machine learning techniques. The Principal Component Analysis (PCA) is considered to reduce the dimension of the data and make it possible to use the Multinomial Logit (MNL) model. SVM and DNN can predict the traffic state using both primary and reduced datasets (ALL and PCA). Moreover, MNL can be used to not only compare the accuracy of models but also estimate their explanatory power. SVM employing primary datasets outperforms other models with the accuracy rate of 79%. Next, the prediction accuracy rates for SVM-PCA, MNL, DNN-PCA, and DNN-ALL are equal to 78%, 73%, 68%, and 67%, respectively. SVM-ALL exhibits better performance in predicting light, heavy, and blockage states, while MNL can predict the semi-heavy state more accurately. Use of the PCA dataset increases the accuracy of DNN and decreases SVM accuracy by 1%. Greater precision is achieved for the rst three months of testing than that in the second three months.