A DATA-DRIVEN FRAMEWORK FOR THE PREDICTION OF NON-RECURRENT TRAFFIC CONGESTION RECOVERY TIME ON FREEWAYS

Abstract

This study introduces a comprehensive approach aimed at improving the management of incident durations. It delves into enhancing traffic incident management by integrating diverse incident datasets, including Maryland State Police incident data and Coordinated Highways Action Response Team (CHART) incident data, to improve the assessment of traffic incident durations. The dissertation employs spatial and temporal thresholds to explore matching different incident datasets and identifies discrepancies between various incident reports. The dissertation also explores methodologies for estimating traffic recovery times of each incident, utilizing historical data and pre-incident conditions as baselines to establish normal traffic conditions. A novel framework is introduced to estimate non-recurrent traffic congestion recovery time, revealing that many incidents recover faster than their reported clearance times. In these cases, traffic flow returns to normal conditions quickly.Further, the study examines predictive modeling for traffic recovery time, highlighting the Random Forest model's effectiveness among various machine learning algorithms. This model's superiority, based on precision, recall, and F1-scores, underlines its potential in accurately predicting traffic incident recovery time categorized as short-duration, medium-duration, and long-duration incidents. In particular, the random forest model results in a precision of 0.7 for short-duration incidents, 0.3 for medium-duration incidents, and 0.5 for long-duration incidents. For instance, the precision of 0.5 for long-duration incidents indicates that half of the cases predicted as long-duration incidents are indeed long-duration incidents. Key predictors such as link-level vehicle volume, clearance time, response time, and number of lanes closed are identified, providing valuable insights for traffic management strategies. This dissertation underscores the importance of data-driven approaches in traffic incident management, aiming to enhance the efficiency of transportation systems through accurate prediction and estimation of incident recovery times.