A Weather Oriented Pre-Tuning Methodology For Long-term Traffic Speed Estimation

Abstract

Long-term traffic speed estimation become a challenging problem since forecasting the distant future of city traffic requires considering environmental factors such as weather, big events, road maintenance, and accidents. One of the predictable factors is weather condition which has a substantial impact on traffic speed, especially in metropolitan cities. It is very important to exploit the weather parameters correctly to predict traffic speed up to 1 week ahead. In this study, we propose to pre-tune the speed data based on weather parameters before feeding it into deep learning algorithms. Two different pre-tuners, Effect Rate (ER) and Polynomial Regression (PR), are introduced where the first method calculates the effect of weather conditions linearly, while the second method proposes to transform the traffic characteristic regarding weather conditions with the help of polynomial regression. Test results showed that the proposed pre-tuners could decrease the traffic prediction error rate up to 20% depending upon the weather condition.