Optimal exploitation of sparsely localised traffic data for modelling traffic flow in road networks

Traffic flow data are generally collected using induction loops. Therefore, to capture traffic dynamics along road sections, a very fine mesh of induction loops is required. However, due to their higher cost of installation and maintenance, induction loops are installed only at some specific areas, namely around road intersections and slip roads. The most optimal exploitation of the sparsely localised data collected via the induction loops installed within a road network, is to combine a prediction model (e.g., a machine learning model), trained using these collected data, with a macroscopic model of traffic flow also known as continuum model. The prediction model enables to forecast the traffic demand and supply downstream and upstream the intersection, respectively, whereas the macroscopic model of traffic flow will be used to capture traffic dynamics, namely the space-time nonlinear features, along the sections of the road non-equipped with induction loops. In this study, we present an application of such approach, where an artificial neural networks model and the well-known Lighthill, Whitham and Richards model (LWR model) are used as prediction and continuum models, respectively. The investigated road network includes parts of the highways A12-A120, A120, A12 and A14 in East Anglia (England, United Kingdom).