The acceleration of urbanization is witnessed all around the world. Both population and vehicle ownership are rapidly growing, and the induced traffic congestion becomes an increasingly pervasive problem in people’s daily life. In 2014, transportation congestion caused $160 billion economic loss in 498 U.S. urban areas, which is 5.5 more than that in 1982. Without effective reactions, this number is expect to grow to $192 billion in 2020. In order to mitigate traffic congestion, many transportation demand management (TDM) strategies (e.g. bus rapid lanes, and flextime policy), and active traffic management (ATM) strategies (e.g. real-time user guidance, and adaptive traffic signal control) have been proposed and implemented.

Although TDM and ATM have proved their values in theoretical researches or field implementations, it is still hard for transportation engineers to select the optimal strategy when faced with complex traffic conditions. In the science of transportation engineering, mathematical models are usually expected to help estimate traffic conditions under different scenarios. There have been a number of models that help transportation engineers make decisions. However, many of them are developed for offline use and are not suitable for real-time applications due to computational time issues. With the development of computational technologies and traffic monitoring systems, online transportation network modeling is getting closer and closer to reality. The objective of this dissertation is to develop a large-scale mesoscopic transportation model which is integrated with an agent-based travel behavior model. The ultimate goal is to achieve online (real-time) simulation to estimate and predict the traffic performance of the entire Washington D.C. area. The simulation system is expected to support real-time transportation system managements and operations.

One of the most challenging issue for this dissertation is the calibration of online simulation models. Model parameters need to be estimated based on real-time traffic data to reflect the reality. Literature review of previous relevant studies indicates a trade-off between computational speed and calibration accuracy. In order to apply the model onto a real-time horizon, experts usually ignore the inherent mechanism of traffic modeling but rely on fast converging technologies to approximate the model parameters. Differently from previous online transportation simulation approaches, the method proposed in this dissertation focuses more on the mechanism of transportation modeling. With the fundamental understanding of the modeling mechanism, one can quickly determine the gradient of model parameters such that the gap between real-time traffic measures and simulation results is minimized.

This research is one of the earliest attempts to introduce both agent-based modeling and gradient-based calibration approach to model real-time large-scale networks. The contribution includes: 1) integrate an agent-based travel behavior model into dynamic transportation network models to enhance the behavior realism; 2) propose a fast online calibration procedure that quickly adjusts model parameters based on real-time traffic data. A number of real-world case studies are illustrated to demonstrate the value of this model for both long-term and real-time applications.