Ensemble Model to Estimate Incident Clearance Durations using Sequential Partitioning Process and Robust Regression

Most agencies, in response to and management of non-recurrent highway congestion, are requested by the general public to provide the estimated delay and impacts of incidents; this information also allows the agencies to take appropriate control strategies. However, to do so in real time the responsible agencies would need to have a reliable estimate of the incident duration, which although valuable information, is either not available or not acceptably reliable for use in practice. Considering the nature of incident response operations, the difficulty in developing a reliable model may be attributed to both the data quality and, most importantly, the many continuous and discrete variables associated with the incident duration; these include resources and staff level of the incident response team, the nature, onset time, and location of the incident, and other related environmental issues. This study therefore proposes an estimation methodology to circumvent these limitations and take advantage of unique characteristics revealed in incident databases for yielding a robust estimate of incident duration. With well-designed partitioning, clustering, and sequential tests to divide all incidents into several distinct groups, the proposed methodology will yield one primary model using all available data and supplemental models for incidents in each group that are calibrated to best fit their unique characteristics and statistical properties. By using incident data from Maryland-CHART, the evaluation results confirm that the proposed methodology can indeed improve the estimation accuracy if properly integrated in the primary model with each supplemental model.