Dynamic ensemble forecasting of traffic flow by means of machine learning techniques

This article presents novel approaches to automatically learn the best combination of forecasts computed by several individual forecast methods. Ideas from the machine learning domain, such as Artificial Neural Networks and Learning Classifier Systems are adapted for this task. The combined forecast serves as basis for a pro-active adaptation of the control strategy in Organic Traffic Control (OTC). OTC is a decentralised, self-organised urban traffic control system that has the ability to optimise the signalisation, to establish progressive signal systems, and to offer route guidance recommendations. Besides analysing the success of the prediction strategy, we demonstrate the positive effect for OTC in terms of a simulation-based evaluation of an urban area situated in Hamburg, Germany. It reflects the actual topology, traffic data from a census, and the actual control strategy performed as reference. As a result, important figures, such as the average waiting times at red lights and the emission values can be decreased significantly. Our findings support the hypothesis that the use of forecasts is beneficial for traffic control