Date of Award

12-2016

Document Type

Thesis

Degree Name

Master of Science (MS)

Legacy Department

Civil Engineering

Committee Member

Mashrur Chowdhury, Committee Chair

Committee Member

Wayne Sarasua

Committee Member

Feng Luo

Abstract

Annual Average Daily Traffic (AADT) is one of the most important traffic parameters used in transportation engineering analysis. Moreover, each state Department of Transportation (DOT) must report the AADT data to Federal Highway Administration (FHWA) annually as part of the Highway Performance Monitoring System (HPMS) requirements. For this reason, state DOTs continually collect AADT data via permanent count stations and short-term counts. In South Carolina, only interstates and primary routes are equipped with permanent count stations. For the majority of the secondary routes, AADT data are estimated based on short-term counts or are simply guesstimated based on their functional classifications. In this study the use of Artificial Neural Network (ANN) and Support Vector Regression (SVR) were applied to estimate AADT from short-term counts. The results were compared to the traditional factor method used by South Carolina Department of Transportation (SCDOT) and also to the Ordinary Least-square Regression method. The comparison between ANN and SVR revealed that SVR functions better than ANN in making AADT estimation for different functional classes. A second comparison was conducted between SVR and the traditional factor method. The comparative analysis revealed that SVR performed better that the traditional factor method. Similarly, the comparison between SVR and regression analysis for the principal arterials revealed no significant difference in the actual AADT and the AADT estimated through SVR. However, it did show a significant difference between the actual AADT and AADT estimated through regression analysis. One of the primary challenges of accurate measurement of AADT is having reliable, complete, and accurate traffic data. Previous research has indicated that transportation agencies often report that a significant portion of their hourly data collected from permanent count stations are missing. This percentage of missing traffic data vary between 10% to 60%. In an effort to address this issue, state departments of transportation either discard missing data or impute the missing data. SCDOT imputes the missing hourly volume using the historical average of the last 3 months’ data from the same day and hour. This method of data imputation could often be erroneous. In order to develop an accurate estimation of missing hourly volume from the permanent count stations, this study applied two machine learning techniques: Artificial Neural Network (ANN) and Support Vector Regression (SVR) for predicting hourly missing data. Data imputation models are developed for Urban Principal Arterial (Interstate), Rural Principal Arterial (Interstate), and Urban Principal Arterials-other functional class. Each of these functional classes were divided into different models based on the on different combination of input features. This study indicated that for each functional class, SVR outperformed ANN. SVR model performance was later compared with current SCDOT’s historic average imputation method, which revealed that SVR model is more accurate in estimating missing values compared to the average imputation method by SCDOT.

Share

COinS