Speed pattern recognition technique for short-term traffic forecasting based on traffic dynamics

摘要

This study introduces a new short-term traffic forecasting technique, based on the dynamic features of traffic data derived from vehicles moving in urban networks. The authors goal is to forecast the values of appropriate traffic status indicators such as average travel time or speed, for one or more time steps in the future until the next half hour. The proposed forecasting technique is based on road profiles generated from the application of data clustering techniques on real traffic data. Data clustering is applied after the original feature space is transformed to a new one of a significantly lower dimension. This transformation is based on the dynamic characteristics of current traffic, expressed in the form of the speed derivatives. To evaluate the proposed technique they used two-week historical data from the city of Berlin, Germany. Extensive evaluation results indicate improvement of the forecasting accuracy after comparison with a set of existing traffic forecasting techniques.