A Scalable Approach for Short-Term Predictions of Link Traffic Flow by Online Association of Clustering Profiles

Short-term prediction of traffic flows is an important topic for any traffic management control room. The large availability of real-time data raises not only the expectations for high accuracy of the forecast methodology, but also the requirements for fast computing performances. The proposed approach is based on a real-time association of the latest data received from a sensor to the representative daily profile of one among the clusters that are built offline based on an historical data set using Affinity Propagation algorithm. High scalability is achieved ignoring spatial correlations among different sensors, and for each of them an independent model is built-up. Therefore, each sensor has its own clusters of profiles with their representatives; during the short-term forecast operation the most similar representative is selected by looking at the last data received in a specified time window and the proposed forecast corresponds to the values of the cluster representative.