Data-Driven Approaches for Traffic State and Emission Estimation

Traffic congestion is one of the most severe problems in modern urban areas. Besides the amplified travel times, traffic congestion intensifies the amount of emitted pollutants impacting human health and the environment. By making the appropriate interventions in traffic, transportation planners can mitigate congestion and enhance the performance of a traffic system. One crucial step in traffic planning and management is the estimation of the current or historical traffic state of a network. The estimation of the traffic state variables (traffic flow, density and speed) reveals the problematic parts of a network, namely, the parts associated with severe congestion and high emission rates. Traffic-related observations and traffic models constitute two core elements of a traffic state estimation approach. While the available observation data explicitly or implicitly provide partial information on the traffic state, traffic models define the traffic behaviour and contribute to estimating the variables when they are not directly observable. The estimated traffic state variables form the input to the so-called emission models, which estimate the mass of the emitted pollutants. The type and availability level of the observation data play a key role in traffic state and emission estimation. Traditionally, the primary source of traffic-related field data are stationary detectors (loop detectors, radar sensors or cameras). Today, following the late advances in communication systems, a vast amount of traffic-related data from mobile sources (GPS or cellular networks) is available. Such high data availability may give transportation planners new insights into understanding traffic behaviour. Appropriate exploitation of data coming from mobile sources can improve the existing approaches for estimating the traffic state and emissions. The broad aim of this thesis is to enhance the quality of traffic state and emission estimation. A special focus is given to the development of methods for exploiting the growing availability of traffic-related field data. By combining traffic data and models, the thesis proposes data-driven approaches for traffic state and emission estimation. The first part of the thesis (Paper I and Paper II) focuses on improving the current approaches for network-wide emission estimation. Traditionally, network-wide emission estimations rely on a static traffic-modelling framework. In Paper I, we suggest an alternative emission estimation approach, which is based on a quasi-dynamic traffic model. To evaluate our approach, we perform field experiments on a 19 km long highway stretch in Stockholm. The results show that our method can improve the spatiotemporal distribution of the estimated emissions. In Paper II, the approach suggested in Paper I is applied to a more extensive network covering the city of Norrköping. The results indicate that our approach yields a realistic spatial layout of emissions. The second part of the thesis (Paper III and Paper IV) suggests novel data-driven approaches for estimating network-wide traffic flows and demand. More specifically, in Paper III, we develop a data-driven traffic-flow propagation approach by utilising traveltime observations. Our method is based on a piecewise linear approximation of the travel time function, which allows the use of an efficient event-based structure for propagating the traffic flow. We evaluate our approach through simulation-based experiments, and the results provide proof of the concept. In Paper IV, we exploit the approach suggested in Paper III to develop an efficient data-driven scheme for estimating the traffic demand. The results of the simulation-based experiments indicate that our approach might lead to more accurate estimations compared to other data-driven estimation approaches suggested in the literature. Finally, the last part of the thesis (Paper V) focuses on the estimation of fuel consumption and emissions at a vehicle level. In paper V, we propose a novel method for generating virtual vehicle trajectories by fusing data from different sources. Our approach provides a detailed description of vehicle kinematics, and thus, it permits the use of the underlying virtual vehicle trajectories to vehicle dynamics-sensitive applications, such as emission modelling. The results of our experiments show that the advanced modelling of vehicle kinematics can enhance the accuracy of the estimated emissions