Short-term prediction of traffic state, statistical approach versus machine learning approach

Abstract

Traffic short-term prediction helps intelligent transportation systems manage future travel demand. The objective of this paper is to predict the traffic state for Karaj to Chaloos, a suburban road in Iran. Two approaches, statistical and machine learning are investigated. We evaluate the performance of the multinomial logit model, the support vector machine, and the deep neural network as two top machine learning techniques. The principal component analysis is used to reduce the dimension of the data and make it possible to use the MNL model. SVM and DNN predict traffic state using both primary and reduced datasets (ALL and PCA). MNL can be used not only to compare the accuracy of models but also to estimate their explanatory power. SVM employing primarily datasets outperforms other models by 79% accuracy. Next, the prediction accuracy for SVM-PCA, MNL, DNN-PCA, and DNN-ALL are equal to 78%, 73%, 68%, and 67%. SVM-ALL has better performance for predicting light, heavy, and blockage states, while the semi-heavy state is predicted more accurately by MNL. Using the PCA dataset increases the accuracy of DNN but decreases SVM accuracy by 1%. More precision is achieved for the first three months of testing compared to the second three months.