Forecasting Traffic Congestion States based on Motorway Grid Cells using Floating Car Data

Congestion is a major and increasingly limiting factor for mobility on motorway networks. Using floating car data, congestion states were usually predicted for road segments that were identified by an additional map-matching tool. Not using a map-matcher, 12 billion floating car data observations were cell-wise grouped into two direction classes representing two directions of a motorway. To the best knowledge of the author, the direction distinctive grid-based approach for assigning floating car data to motorway segments is proposed for the first time in this study. The well-known random forest classification algorithm was utilised for developing and forecasting models for single segments and the segment’s collective of 1,000 motorway segments and 45 million observations. Evaluation was based on the metrics F1-score, misclassification rate, and Bookmaker Informedness. Heuristics based on the average velocity of all motorists at specific points on the motorways served as ground truth for forecasting a segment’s congestion state into one of the two classes: free-flowing and congestion. Whole grid forecasting models delivered better results in comparison to single segment models for four highly congested motorway segments. Major influential factors for the five-minute forecast of the segment collective were features regarding the velocity and the traffic count. Whole grid models are seemingly capable of adding value to the congestion state forecast in 5, 10, 20, 30, and 60 minutes in the future in the whole of North Rhine-Westphalia by considerably exceeding the F1 and Bookmaker Informedness baseline scores for the 1,000 sample segments. The computational effort was more than 30% lower when using the direction distinctive grid-based approach in comparison to a map-matcher approach for assigning road segments to 2.6 million floating car data observations