Predictive distance-based road pricing — Designing tolling zones through unsupervised learning

Congestion pricing is a standard approach to mitigate traffic congestion in a number of urban networks around the world. The advancement of satellite technology has spurred interest in distance-based congestion pricing schemes, which obviate the need for fixed infrastructure such as gantries that are used in area- and cordon-based pricing. Moreover, distance-based pricing has the potential to more effectively manage traffic congestion. In the context of distance-based congestion pricing, we propose the use of sparse subspace clustering methods employing Elastic Net optimization (SSCEL) and Orthogonal Matching Pursuit (SSCOMP), as well as two hierarchical density-based clustering methods, (OPTICS, HDBSCAN*) for the derivation of tolling zones. These tolling zones are then used within a simulation-based framework for real-time predictive distance-based toll optimization to examine network congestion and performance of the tolling schemes. Within this framework, for a given definition of tolling zones, tolling function parameters are optimized in real-time using a simulation-based Dynamic Traffic Assignment (DTA) model. Guidance information generation is integrated into the predictive optimization framework and behavioral responses to the information and tolls along dimensions of departure time, route, mode, and trip cancellation are explicitly modeled. For the evaluation of network performance we make use of Travel Speed Index (TSI) data from the real-world Boston Central Business District urban network and demonstrate that tolling zones derived from the sparse subspace clustering are an effective means of operationalizing real-time distance-based toll optimization schemes, showing improvements in average travel time and social welfare relative to the baseline.