Decision Support System for Predicting Traffic Diversion Impact Across Transportation Networks Using Support Vector Regression

Document Type

Journal Article

Publication Date


Subject Area

operations - traffic, land use - impacts, ridership - commuting, ridership - forecasting, ridership - forecasting, organisation - management


Vermont, Transportation networks, Traffic incidents, Traffic diversion, Support vector regression, South Carolina, Scenarios, Projections, Interstate highways, Incident management, Impacts, Icon based interfaces, GUI, Graphical user interfaces, Forecasting, Diverted traffic, Decision support systems, Case based reasoning


This paper describes follow-up research to a previous study by the authors that used case-based reasoning (CBR) and support vector regression (SVR) to evaluate the likely impacts of implementing diversion strategies in response to incidents on highway networks. In the previous study, the training and testing of the CBR and SVR tools were performed on a single transportation network from South Carolina, which limited the applicability of the developed tool to the specific network for which it was developed. To address this limitation, the current study investigates the feasibility of developing a generic decision support system (DSS) capable of predicting traffic diversion impacts for new transportation networks that the tool has not previously seen. In such cases, users need only to input the geometric and traffic variables, via a graphical user interface, and the tool, which uses a SVR model, will predict the benefits of diverting traffic for a specific incident on the new site. To illustrate the feasibility of developing such a tool, two different highway networks covering portions of I-85 and I-385 in South Carolina were used to train the SVR model, which was then tested on a third network covering portions of I-89 in Vermont. The study found only a 15% difference between the predictions of the SVR model and those of a detailed simulation counterpart, demonstrating the feasibility of developing a generic DSS. Adding more sites and parameters to train the software is also expected to improve the prediction accuracy of the DSS.