Forecasting with Data-Driven Models
Corresponding Author: Rick Donnelly, WSP
Presented By: Rick Donnelly, WSP
Abstract
Several instances of travel models built using large-scale passively collected cellular location data have started appearing in the literature. Such models are typically built in combination with other public and commercial databases, rather than with travel diaries of a small percentage of the local population. Examples include several truck trip matrices derived from GPS tracking data and a tour-based model built by the authors entirely from passive data. One perceived limitation of such models is that they cannot be used in forecasting, for the type of data used to build them are not available from the future. This presentation will focus upon how these data- and pattern-driven models can in fact be used for forecasting. First, a review will be conducted on how traditional models incorporate expected changes in inputs and behavior. In parallel the same will be outlined for how such changes can be incorporated into data-driven models. Concrete examples will show how changes in networks and populations can be handled, as well as how assumed changes in trip generation, destination choice, and traffic assignment can be uniquely represented in such models. A conceptual framework for handling mode choice will also be presented. The emphasis will be upon techniques that can be applied now, rather than expected future advances.