Spatio-temporal prediction of freeway congestion patterns using discrete choice methods

Predicting freeway traffic states is, so far, based on predicting speeds or traffic volumes with various methodological approaches ranging from statistical modeling to deep learning. Traffic on freeways, however, follows specific patterns in space–time, such as stop-and-go waves or mega jams. These patterns are informative because they propagate in space–time in different ways, e.g., stop and go waves exhibit a typical propagation that can range far ahead in time. If these patterns and their propagation become predictable, this information can improve and enrich traffic state prediction. In this paper, we use a rich data set of congestion patterns on the A9 freeway in Germany near Munich to develop a mixed logit model to predict the probability and then spatio-temporally map the congestion patterns by analyzing the results. As explanatory variables, we use variables characterizing the layout of the freeway and variables describing the presence of previous congestion patterns. We find that a mixed logit model significantly improves the prediction of congestion patterns compared to the prediction of congestion with the average presence of the patterns at a given location or time.