Traffic flow prediction is widely used in travel decision making, traffic control, roadway system planning, which is extremely necessary for individual travelers, business sectors, and government agencies. ARX models have proved to be highly effective and versatile. In this research, we investigated the applications of ARX models in prediction for real traffic flow in New York City. The ARX models were constructed by linear/polynomial or neural networks. For linear/polynomial networks, Least angle regression (LARs) method was firstly implemented to determine appropriate features. Ordinary least square (OLS), ridge regression (RR), and least absolute shrinkage and selection operator (Lasso) were then applied to determine model coefficients. For neural networks, we considered both shall recurrent neural network (SRNN) and long short term memory (LSTM), which were trained by Levenberg-Marquardt (LM) and adaptive moment estimation (Adam) algorithms, respectively. The exogenous inputs includes traffic volume data for neighbour roads and weather data. Comparative studies were carried out based on the results by the efficiency, accuracy and training computational demand of the algorithms.

CICTP