Modeling Framework for Predicting Lane Change Intensity at Freeway Weaving Segments

Abstract

This study proposes a modeling framework for predicting discretionary lane change (DLC) intensity at weaving segments using traffic flow and site data. The database used to develop the models comprises 294 field observations from 19 sites. Two modeling techniques, using regression trees and linear regression, were employed to predict DLCs per hour and DLCs per vehicle. The proposed models were compared with the lane change model for weaving segments in the Highway Capacity Manual (HCM7). The lane change data were clustered by site, which cautioned the applicability of linear regression for this dataset. Nonetheless, both the regression tree and linear regression models yielded high R-squared values, varying from 0.93 to 0.96. The relative root mean squared error (RMSE)—the ratio of the error to the mean values—varied between 0.18 and 0.30. However, a site-specific validation showed that the linear regression models performed poorly for most sites, although measures were taken to cope with outliers, nonlinearity, and interactions. The tree model improved the prediction of DLCs per hour for more than two-thirds of the sites when compared with the mean value at each site. It also performed well in most cases when applied to a site that was omitted from the model development. The HCM7 model performed well when applied to an omitted site. However, it exhibited the highest overall relative RMSE (0.57), underscoring the necessity of advanced modeling tools with additional predictors. We recommend incorporating observations from more extended periods and varying traffic conditions for each site for future research.