Control Oriented Freeway Traffic Modelling by Physics-Regularized Machine Learning

A modelling framework for freeway traffic systems specifically oriented to the definition of suitable control schemes is developed in this paper. The traffic system includes different vehicle classes (passengers and freight vehicles) whose behaviour is explicitly modelled. Vehicles can enter the free-way arriving from the mainstream or through on-ramps on which traffic volumes may be regulated with ramp metering techniques. Moreover, Artificial Intelligence (AI) techniques and, specifically a Machine Learning (ML) approach, are the modelling methodology proposed in this work. More in detail, a hybrid model is defined in which a physics-based component, provided by a multi-class version of METANET model, is added to a machine learning model as a regularization component and the derived model is trained to reproduce the behavior of different classes of users. The effectiveness of the hybrid model is shown by training and testing it with real traffic data.

I. Introduction

The adoption of effective traffic control schemes is a crucial need, since nowadays both urban and extraurban arterials are still affected by recurrent and non-recurrent congestion phenomena, due to a massive and always increasing use of road transport. The social, economic and environmental impact of road congestion is very high, thus representing a barrier for a smart, safe and clean development of countries. A suitable approach to tackle congestion is the design of effective management and control techniques aimed at effectively exploiting the existing infrastructures.