Application of Unsupervised Machine Learning to Increase Safety and Mobility on Roadways after Snowstorms

摘要

The impact of a snowstorm on the safety and mobility of roadway transportation depends mainly on the storm's level of severity. Defining storms' severity, though, is challenging due to the high number of weather parameters needed to describe these events and the non-linear relationships among these parameters. Finding patterns among snowstorms can conceivably simplify this process and help practitioners better analyze and prepare for such events, even when the severity is not explicitly quantified. Therefore, this study interrogated historical data to assess and compare clustering methods and to identify patterns manifesting in snowstorms to lay the necessary foundations for building a more reliable and objective winter severity index. The research team selected three hierarchical clustering methods that differentiated similar groups of snowstorms among more than 2,000 events dated between 2006—2016 in Nebraska. The team then evaluated the performance of these methods using the Calinski-Harabasz index. A range of clustering scenarios were reviewed visually using principal component analysis to determine the optimal number of clusters. The results indicate that while some districts can be described by as few as three clusters, others can experience up to six different clusters of snowstorms. The use of PCA and visualization in this context can facilitate a better understanding of these high-dimensional data, and the findings of this study can help agencies better comprehend snowstorms and prepare for them, which can help communities to maintain the safety and mobility of their drivers.