Knowledge-Based System for Estimating Incident Clearance Duration for Maryland I-95

Abstract

For incident response operations to be appreciated by the general public, it is essential that responsible highway agencies are capable of providing the estimated clearance duration of a detected incident at a level sufficiently reliable for motorists to make proper decisions such as selecting a detour route. Depending on the estimated clearance duration, the incident response center can then implement proper strategies to interact with motorists, ranging from providing incident information only to executing mandatory detouring operations. This study presents a knowledge-based system, based on detailed incident reports collected by the Maryland-CHART (Coordinated Highway Action Response Team) program between years 2012 and 2016, for such needs. The proposed system features the use of interval-based estimates derived from knowledge of historical data, with different confidence levels for each estimated incident clearance duration, and its rule-based structure for convenient updates with new data and available expertise from field operators. As some key variables associated with incident duration often only become available as the clearance operations progress, the developed system with its sequential nature allows users to dynamically revise the estimated duration when additional data have been reported. The preliminary evaluation results have shown the promise of the developed system which, with its invaluable historical information, can circumvent the many data quality and availability issues which have long plagued the applicability of some state-of-the-art models on this subject.

Outlier Analysis to Improve the Performance of an Incident Duration Estimation and Incident Management System

“…Each of those models is the only best fit for some datasets with unique characteristics or properties. In view of many qualitative and quantitative variables associated with incident durations and those complex correlations ( 23 ), it would be difficult to use only one of the existing outlier detection algorithms for various data patterns.…”

“… G T and G V denote the training and validation sets of an outlier data group, respectively; O is the selected outlier from each dataset; i is the number of selected outliers with the longest duration; therefore, O i = 1 shows one outlier with the longest duration and O i = 2 denotes two outliers with the first and second longest duration, and so on. A key feature of the proposed approach with the hybrid association rule mining method is that one can preferentially identify the pattern of outliers with a long duration in the highly imbalanced classification problem ( 23 ). Details of this proposed approach to recognize patterns of the outliers are described in the author’s previous studies ( 23 , 24 ).…”

“…A key feature of the proposed approach with the hybrid association rule mining method is that one can preferentially identify the pattern of outliers with a long duration in the highly imbalanced classification problem ( 23 ). Details of this proposed approach to recognize patterns of the outliers are described in the author’s previous studies ( 23 , 24 ).…”

“…Many studies have attempted to develop robust and reliable incident duration prediction models, which range from continuous models to hazard-based models ( 2 – 10 ), including quantile regression ( 11 , 12 ), rule-based ( 13 , 14 ), and hybrid ( 15 – 24 ) models. Despite these considerable efforts, however, most existing incident datasets share the following common issues that need to be addressed: (i) a highly right-skewed distribution, (ii) many qualitative variables, (iii) complex relations among the key contributing factors, (iv) multicollinearity, (v) heteroscedasticity, (vi) faulty and missing information in the data; and (vii) outliers.…”

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

“…In reviewing the literature, it is noticeable that there exists a large body of studies for estimating incident duration, ranging from continuous models such as hazard-based models ( 1 – 9 ) and quantile regression models ( 10 , 11 ), to discrete/rule-based models such as regression tree models ( 12 , 13 ) and hybrid models ( 14 – 22 ). For example, to deal with the unique characteristic distribution of incident duration data, some researchers developed hazard-based models for estimating incident duration ( 1 – 9 ).…”