Ensemble Learning Approach for Freeway Short-Term Traffic Flow Prediction

Abstract

As traffic situations deteriorate in metropolitan areas around the world, intelligent transportation systems (ITS) emerge as a promising technology. One key issue in the ITS is the problem of short-term traffic flow forecasting which targets at forecasting traffic flow value in the near future (short-term) based on the real time data and historic data collected by data gathering systems in transportation networks. A lot of approaches have been proposed in past references to forecast short-term traffic flow. Time-series-based method, Kalman Filter method, nonparametric method and neural-networks-based method are representative approaches. However, although researchers have proposed those prediction methods and declared their validities and efficiencies, no one has devoted on improving prediction capabilities through ensemble learning methods continuously. This paper explores how the ensemble learning method, namely bagging, remarkably decreases the prediction error such as in the radial basis function neural network. Moreover, real freeway short-term traffic flow predictions such as the effects of the extent of prediction, the "look-back" interval and the time resolution on the prediction accuracy are carefully studied based on a real traffic flow data gathered at Loop 3 freeway in Beijing, China.