Modeling Incident Duration Using Lasso and Ridge Regressions

Abstract

Traffic incidents significantly disrupt freeway operations, causing delays, congestion, fuel waste, and economic losses. Effective incident management requires not only rapid detection and clearance but also accurate real-time prediction of incident duration, a capability currently lacking in most Traffic Management Centers (TMCs). This study develops machine learning models to predict incident duration based on real-time responses and evolving incident conditions. The analysis uses incident data from the I-15 corridor in Las Vegas, Nevada, encompassing 643 recorded incidents, of which 272 were further documented in a novel Video Snapshots Dataset (VSDS) that captures 15-second visual records of incident characteristics. Key incident attributes, including total and average blockage duration, were extracted to enrich model training. Three predictive approaches were evaluated: Lasso Regression and Ridge Regression with 10-fold cross-validation, with comparison with Multiple Linear Regression. Among these, Ridge Regression achieved the highest predictive accuracy, demonstrating its effectiveness for real-time estimation of incident duration. These findings provide a foundation for enhancing TMC operations by enabling more reliable travel time updates and proactive traffic management strategies.