Using Real-Time Weather Information for Traveler Information: A Statistical Learning Application UnderA lternative Experimental Conditions

摘要

This paper examines the case where information on real-time weather conditions is used to predict future' speeds in a traveler's route so that they can make travel decisions relating to whether or not to make the trip, change departure time or to take an alternative route, if already on the road. We examine the performance of two different classes of models (a "base" model and a "statistical learning" model) to predict future speeds while controlling for location, demand/time-of-day, prevailing speeds, and future weather conditions. A stratified sampling strategy is adopted with good and bad weather conditions sampled separately within locations along the Eisenhower Expressway in Chicago. We find differences in the predictive abilities of the two models under different weather conditions. The SVM model outperforms the linear regression model by predicting 41% more cases within 3mph of the observed speed under heavy rain conditions, 16.7% more cases within 3mph under snow conditions, 14.5% more cases within 3mph under thunderstorm conditions, and with less dramatic differences under other weather conditions. The results show promise that SVM regression maybe useful in bringing together streaming forecasted weather data and traffic conditions to inform travelers.