Integrated Traffic State Uncertainty Modeling, Portable Traffic Sensor Network Planning, and Management

Abstract

Over the past few years, traffic congestion has become a genuine nightmare to most of the urban commuters. Providing real-time traffic information is of key significance to Intelligent Transportation System (ITS). Accurate travel time or traffic speed information through Advanced Traveler Information System (ATIS) can provide guidance for travelers who make decisions every day on travel mode, route choice and departure time. Meanwhile, travelers' anxiety can be reduced with a better understanding of their current and future travel time. With a well-organized and reliable traffic surveillance network, ITS can not only assist travelers in understanding their travel time and planning their trips via ATIS but also detect traffic incident and dispatch a patrol team in a timely manner. Therefore, comprehensive and reliable traffic network surveillance is of fundamental significance in building a smart transportation network. This dissertation deals with three major issues about the highway system. Traffic state such as travel time or traffic speed serves as a key parameter to reflect the highway system operation efficiency. Understanding the real-time traffic information is useful in helping travelers make smart route choice and schedule proper departure times. Lots of efforts have been made to improve traffic state prediction performance with advanced real-time prediction models. But there is limited work studying the intrinsic prediction uncertainty of such data-driven based predictions. This dissertation developed an entropy-based uncertainty estimation model to evaluate system state predictability under any given measurement space from a stochastic evolution perspective. Then we considered the highway network as a stochastic system and applied the proposed model to evaluate travel time prediction uncertainty under both temporal and spatial measurement spaces. Moreover, the quantitative relationships between data-driven based prediction errors and the proposed uncertainty measurements are analyzed based on a real-world case study.

Second, we developed a sensor network optimization model aiming to provide network-level real-time traffic information surveillance. The proposed model has two advantages compared with traditional traffic sensor planning models. Conventionally, people only focus on the surveillance benefit at the location where sensors are placed while ignoring the surveillance benefit improvement inferred from the spatial traffic state correlations. Moreover, the proposed network optimization model provides one with the flexibility to come up with optimal sensor relocation strategies. Specifically, when traffic demand and travel time uncertainty are heterogeneously distributed in a highway network for a given time period, appropriately relocating sensors can fully make use of the surveillance resources and enhance the network surveillance. The proposed model was applied to plan a travel time surveillance network for Washington D.C.-Baltimore commute network. Optimal sensor placement strategies and relocation operations with respect to the surveillance benefits are analyzed and discussed for the study area.

Last, we consider the sensor placement problem from a different perspective given the a priori information is completely missing. For a highway network with complete unknown historical traffic data and unknown GPS coverage, the question that how operators should plan a sensor network to evaluate these a priori traffic information is answered. Specifically, a multistage stochastic optimization model with endogenous uncertainty is presented, and a Monte Carlo simulation-based approach is designed to evaluate the optimal solution. The proposed optimization model was applied to the same Washington D.C.-Baltimore commute network and serves as a supplement to the real-time surveillance based dynamic sensor network model.