Using Probe Data Analytics For Characterizing Speed Reductions as well as Predicting Speeds During Rain Events
2020-07-29T13:44:39Z (GMT) by
This study emphasizes the extreme variability present in traffic speed studies and the need for high resolution traffic and weather data in order to understand the interaction between traffic speeds and weather. I analyzed the impact rainfall has on roadway traffic speeds along I-65 in Indiana for the month of June 2018 and attempted to leverage this information to model and predict traffic speeds. To develop a statistical distributional understanding of the difference between traffic speeds under rain and non-rain conditions, Quantile-Quantile plots were generated in addition to fitting both scenarios to a gamma distribution. To compare how traffic speeds react to various precipitation intensities, boxplots were generated for comparison. Then, a baseline speed was defined using the median traffic speed under non-rain scenarios and was used to calculate speed reductions from the baseline at varying precipitation intensities. Finally, an XGBoost model is developed to attempt traffic speed predictions. There are five key findings indicated by this study. First, the non-rain traffic speeds above the 5th percentile are typically faster than their rain speed counterparts at comparable quantile levels. Second, traffic speeds exhibit a high amount of variance at varying precipitation intensity levels. Third, the gamma distribution does not suit traffic speed distributions at all locations and times of day under rain or non-rain scenarios. This result is consistent with previous findings that suggest traffic speed interactions are highly variable and based on a variety of factors that are hard to account for. Fourth, weekday traffic speeds from 1600 to 2200 UTC are the most strongly impacted across all regions during rain events seeing speed reductions of up to 10 mph, this is consistent with previous findings. Finally, the XGBoost model did not perform adequately in the configuration used in this study. The poor performance of the XGBoost model was somewhat anticipated as this study did not have access to traffic volume information and instead leverages proxy variables to account for this. The findings of this study demonstrate the need for finer scale studies on traffic—weather interactions and provides methodology that can be extended to other weather and traffic datasets.