FORECASTING TRAVEL TIME AND VARIATIONS IN TRAVEL TIME DUE TO VEHICLE ACCIDENTS IN SPATIO-TEMPORAL CONTEXT ALONG FREEWAY
Analytics
325 views ◎270 downloads ⇓
Abstract
The growth in population, technological advancements, increase in average life expectancy, and newer and fuel-efficient vehicles has led to a meteoric rise in travel demand over the past few decades. However, the road network capacity has not increased at the same progressive rate, resulting in congestion and associated traffic problems. Traffic incidents are major contributors of non-recurring congestion in most of the urban areas in the United States.Travel time is an effective parameter to quantify the congestion at segment- or corridor-level. Short-term traffic and travel time prediction plays a vital role in advanced traveler information systems (ATIS) and assists in proactive management of transportation network. However, forecasting travel time is complex under over-saturated conditions and in the presence of an incident. Besides, travel time itself cannot explain the exact impact of the incident as it varies with respect to time and over space. Incorporating all factors that affect traffic and travel time increases the magnitude of the complexity. Therefore, this research focuses on an application of Autoregressive Integrated Moving Average (ARIMA) model, incorporating travel time information over space (from neighboring segments) and time, to forecast travel time and relative variations in travel time (RVTT) along a freeway corridor in spatio-temporal context. The RVTT was considered instead of variation in travel time to negate the effect of difference in segment lengths and other geometric characteristics for a meaningful comparison. Two types of RVTT were considered: travel time / expected travel time and travel time / minimum travel time.Travel time data was collected from INRIX and incident data was gathered from Traveler Information Management System (TIMS) from 2010 to 2012. Databases were developed using data, for 150 "vehicle accident" affected days and 100 sample days of data when there were no incidents, along a ~19-mile freeway corridor of I-77 S in the city of Charlotte, North Carolina.Four categories of Cronbach’s α were computed at 10-minute intervals for each segment. The higher value was selected as the corresponding Cronbach’s α to capture the expected travel time of that segment. Minimum travel time of the segment was estimated as the minimum of all the travel time samples in a 10-min interval for each day-of-the-week. However, when the minimum travel time was zero then the second minimum travel time was taken as it was assumed that no vehicle passed that segment or data was recorded during the specific time interval.Results obtained indicate that the difference between the observed and the expected travel time is less than 10% for almost 85% samples and less than 15% for 90% of the samples considered from 2010 to 2012. This indicates the effectiveness of Cronbach’s α in capturing the expected travel time for a certain time interval of a segment. Lagged regression model was then developed using data for 18 segments. Three scenarios (1) travel time (TT), (2) travel time/expected travel time (TT/ExpTT), and (3) travel time/ minimum travel time (TT/MinTT) were considered for both without incident and under "vehicle accident" conditions. Developed models from all three scenarios showed that the travel times and RVTT for consecutive segments are highly correlated, and both upstream and downstream segments have an influence on the current state of the target segment. Moreover, all the significant predictor variables had a time lag of 10 minutes. In other words, the prediction horizon is 10 minutes.The Mean Absolute Percent Error (MAPE) and Mean Absolute Deviation (MAD) of the developed lagged regression model are estimated for every segment, irrespective of the incident condition. MAPE and MAD values of all segments of TT are less than 10% for all but one segment, for which MAPE value was marginally greater than 10%. For both TT/ExpTT and TT/MinTT scenario, MAPE and MAD value for all segments was less than 10%, except for one segment.For model validation, a total of 80 days of data was considered (45 without incident days and 35 "vehicle accident" affected days). Results showed that MAPE and MAD values are less than 15% for TT and TT/MinTT scenarios of almost all the segments. However, for TT/ExpTT scenario, all of them are less than 15%. Moreover, both the calibrated and validated models demonstrated that modeling using TT/ExpTT would yield accurate results than TT/MinTT. Except for four segments, forecasting accuracy of TT/ExpTT was higher than TT/MinTT for all other segments. Overall, the adopted methodology successfully forecasted travel time and variations in travel time for both without incident and under "vehicle accident" condition.