Robust ensemble method for short-term traffic flow prediction

Accurate and real-time prediction of short-term traffic flow plays an increasingly vital role in the successful deployment of Intelligent Transportation Systems. Although existing studies have been done for traffic flow prediction problem, their efficacy relies heavily on traffic data. However, collected traffic data are usually affected by various external factors (e.g, weather, traffic jams and accidents), leading to errors and missing data. This makes it difficult to pick a single method that works well all the time. This paper concentrates on investigating ensemble learning that benefits from multiple base methods and presents an effective and robust ensemble method by using the bagging ensemble technique averaging to improve the traffic flow prediction performance. To enhance the robustness of constructed ensemble method, three improved least squares twin support vector regression methods are proposed based on robust L1-norm, L2,p-norm and Lp-norm distance to alleviate the negative effect of traffic data with outliers. In addition, a pruning scheme is utilized to remove anomalous individual components. This makes the proposed method more effective for traffic flow prediction. Further, a comprehensive traffic flow indicator system based on speed, traffic volume, occupancy and ample degree is utilized to forecast the traffic flow. To promote the prediction performance, we optimize the parameters of each component in ensemble method with the adaptive particle swarm optimization. The results on real traffic data demonstrate that the proposed ensemble method yields better prediction performance and robustness even when the standalone components and other competitors make unsatisfactory predictions.