Synergy of Model-driven and Data-driven Approaches in a Dynamic Network Loading Problem

Modern dynamic models of traffic flow and especially dynamic network loading (DNL) models are a powerful approach to predict traffic flow dynamics in a short-term sense (minutes or hours ahead). Such models should be the core element of any intelligent transportation system to make safer and smarter use of transport networks. Nowadays a variety of traffic data is becoming more and more accurate and available. Online traffic data can be incorporated in DNL model to take into account nonrecurring events (e.g. accidents, road closures or unexpected bad weather conditions). This idea can increase the accuracy of short-term prediction and make traffic flow management more effective. In our research we suggest to combine traditional model-driven approach with a data-driven prediction. As a DNL model we use the link transmission model in cooperation with a dynamic user equilibrium algorithm to identify the routes. Traffic data are the values of speed and flow with a 5-minutes time step, obtained from stationary road sensors. We use the rolling horizon approach, that is, every 5-minutes model constructs 1-hour forecast incorporating actual sensor data. Moreover, we use methods of machine learning to predict the sensor data for the next hour and take it into account while calculating the forecast for the current hour ahead.