Systematically Assessing and Comparing Data-Driven Traffic Speed Prediction Models After Crashes

Abstract: Traffic crashes, ranging from minor incidents to severe fatalities, present significant challenges to both safety and mobility on various road types, including freeways and arterials. Accurate prediction of traffic speeds following crashes of different severities is essential for mitigating congestion, preventing secondary crashes, and enhancing traffic management strategies. This study systematically examines the factors influencing post-crash speeds by analyzing crashes of varying severities across both freeways and arterial roads, followed by an evaluation of the capabilities of different machine learning algorithms in predicting these post-crash speeds.
Using high-resolution INRIX probe-vehicle speed data provided by the Arizona Department of Transportation, we investigate how crash severity, road type, and other contextual factors affect speed recovery patterns. Additionally, we compare the performance of various machine learning models – categorized into parametric (e.g., Multiple Linear Regression) and non-parametric (e.g., Random Forest Regression) – in predicting post-crash speeds. Each model is trained and tested using crash data, with prediction accuracy evaluated through four common performance metrics. Results indicate that non-parametric models, particularly Random Forest, outperform parametric models, providing superior predictive accuracy and computational efficiency.
These findings equip transportation agencies and first responders with improved tools for real-time traffic management and forecasting in the aftermath of crashes, contributing to safer and more efficient roadway operations.