Travel time forecasting from clustered time series via optimal fusion strategy

This paper addresses the problem of travel time forecasting within a highway. Several measurements are captured describing travel times for multiple origin-destination (OD) pairs. A network model is then proposed to infer travel time between origin and destination based on a reduced number of states. The forecast strategy is based on current day and historical data. Historical data is organized into several clusters. For each cluster, a predictor is designed based on the Kalman filtering strategy. Then these predictions are fused, in a best linear unbiased estimation sense, in order to get the best prediction. The performance of the proposed method is evaluated using traffic data from the South Ring of the Grenoble city in France.

I. Introduction

Traffic forecasting is one of the most desired tools for traffic management requested by operators and commuters. In the era of data deluge in which we are, measurements collected by sensors are important sources of information that require analysis, classification, and processing in order to detect patterns and behaviours that can be exploited for traffic prediction [7], [14]. The collected information can be classified by clusterization algorithms, such as K-means, where each cluster collects traffic patterns which in some cases characterize typical regimes such as congestion. Several indicators like travel time, queue length, density, delay are used as performance indexes to determine the status of a traffic network [16].