Traffic modeling, forecasting and assignment in large-scale networks

Today, the development and evaluation of traffic management strategies heavily relies on microscopic traffic simulation models. In case detailed input (i.e. od matrix, signal timings, etc.) is extracted and incorporated in these simulators, they can provide valuable traffic state predictions. However, as this type of information is almost never available at the large-scale and traffic represents chaotic behavior in saturated networks, microscopic simulation models remain intractable and unstable. An alternative is a recently discovered network traffic model; macroscopic fundamental diagram (MFD). Nevertheless, large-scale traffic management strategies remain a big challenge partly due to unpredictability of choices of travelers (e.g. route, departure time and mode choice). Part I of the thesis is an attempt to fill this gap. Chapter 2, 3 and 4 elaborate new aspects of large-scale traffic modeling, and integrate route choice behavior into the modeling. Chapter 2 proposes a dynamic traffic assignment (DTA) model to establish equilibrium conditions in multi-region urban networks where the modeling is done through MFD dynamics. The method handles the stochastic components of the aggregated model through a sampling approach. In addition, the assignment model enables us to consider the response of drivers to changing traffic conditions in an aggregated modeling framework. Chapter 3 extends the DTA model presented in Chapter 2 to a route guidance system, where drivers are given a sequence of subregions to follow. Two aggregated models, region- and subregion-based models, are introduced to develop the guidance scheme and to test its effect, respectively. Notably, the challenge here is to translate certain variables across the traffic models without a loss of significance and assure certain degree of consistency. Chapter 4 extracts and reconstructs aggregated route choice patterns through an extensive GPS data set from taxis in a mega city. Observed GPS trajectories are first grouped together to provide a physical evidence for consistent route patterns. Second, in order to investigate the consistency of equilibrium assumptions considered in Chapter 2, observed trajectories are replaced with shortest path trajectories, and aggregated route choice patterns are reconstructed. Part II introduces novel travel time prediction and variability models. Travel time is a crucial performance measure in assessing the efficiency of transportation systems, and it provides a common index for both practitioners and travelers. Chapter 5 develops a travel time prediction model that jointly exploits traffic flow fundamentals and advanced data mining techniques. The prediction method detects the congestion patterns through the identification of active bottlenecks, and clusters the days with similar traffic patterns. This approach basically allows the model to train its predictions with relevant historical data sets. The method is applicable in oversaturated conditions and consistent with physics of traffic flow. Nevertheless, travelers not only consider travel time on average, but also value its variation. Day-to-day travel time variability, addressing the travel time variations of vehicles crossing the same route at the same period of time on different days, reveals interesting patterns. Departure time periods with similar mean travel times in the onset and offset of congestion exhibit quite different variance values. This phenomenon causes counter-clockwise hysteresis loops on the mean-variance curves. Chapter 6 investigates the empirical implications of hysteresis shape within the context of day-to-day travel time variability.