1 Incident Detection and Incident-Impacted Traffic Prediction for Urban Road Networks, Application to GrandLyon Laura Wynter, Sebastien Blandin IBM Research Singapore Collaboratory, Singapore lwynter@sg.ibm.com Barry M Trager, Yichong Yu IBM Research, Yorktown Heights, NY USA Yiannis Kararianakis Arizona State University, Tempe, Arizona, USA Jean Coldefy, Dimitri Marquois ITS and Mobility services unit, GrandLyon, Lyon, France Thomas Baudel IBM France Lab & Efficacity Institute, France ABSTRACT This paper describes joint work done by IBM Research (development of the solution) and GrandLyon (assessment of the solution) for real-time incident detection and traffic prediction on an urban road network in the presence of incidents. The detection and prediction methods are integrated and tested in a pilot study on live real-time traffic data in Lyon Metropolis, France. Numerical results and analysis are provided. The benefits of faster and more accurate incident detection include faster incident clearing times and more accurate and timely information to the public so as to avoid affected areas on the road network. 1. INTRODUCTION Real-time traffic prediction is a critical component of modern road traffic management systems. Indeed, for most traffic management functions to perform adequately, the raw real-time traffic data is obsolete by the time it is received. Hence, real-time traffic prediction, which forecasts the future traffic state from a few minutes to one hour into the future, using real-time data, is needed for intelligent transportation systems applications, such as real-time route guidance, adaptive traffic control and advanced traveler information systems. While traffic prediction technology is highly effective in most network conditions, when an incident occurs on the road network, the statistical traffic prediction methods employed by spatiotemporal time series prediction methods such as those in Min and Wynter (2011) and Kamarianakis, Shen and Wynter (2012) may not perform adequately, especially at the beginning of the incident. Indeed, parameters used by time series methods are calibrated on historical conditions, which include only a limited number of incidents. However, the number of incidents occurring on the road network at any particular place and time is relatively low. We therefore propose a different approach comprised of an