Regime-Based Short-Term Multivariate Traffic Condition Forecasting Algorithm

Abstract

Predictions of fundamental traffic variables in the short-term or near-term future are vital for any successful dynamic traffic management application. Univariate short-term traffic flow prediction algorithms are popular in literature. However, to facilitate the operationalities of advanced adaptive traffic management systems, there is a necessity of developing multivariate traffic condition prediction algorithms. A new multivariate short-term traffic flow and speed prediction methodology is proposed in this paper where the traffic flow and speed observations from uncongested (or linear) and congested (or nonlinear) regimes are regime-adjusted to ensure consistent system dynamics. The prediction methodology is developed by using artificial neural networks (ANN) algorithms in conjunction with adaptive learning rules. These learning rules demonstrate significantly improved accuracy and simultaneous reduction in computation times. Additionally, the paper attempts to identify the most suitable adaptive learning rule from a chosen pool of rules. The validation of the prediction methodology is performed by using traffic data from multiple locations in the United Kingdom (U.K.). The results indicate that the proposed multivariate forecasting algorithm is effective and computationally parsimonious to simultaneously predict traffic flow and speed in freeway or highway networks.