Date of Graduation

12-2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Civil Engineering

Advisor/Mentor

Hernandez, Sarah V.

Committee Member

Hall, Kevin D.

Second Committee Member

Williams, Brent D.

Third Committee Member

Hyun, Kate

Keywords

Freight transportation; Global Positioning System (GPS); Machine learning; Pattern recognition systems; Trip chaining

Abstract

To construct, operate, and maintain a transportation system that supports the efficient movement of freight, transportation agencies must understand economic drivers of freight flow. This is a challenge since freight movement data available to transportation agencies is typically void of commodity and industry information, factors that tie freight movements to underlying economic conditions. With recent advances in the resolution and availability of big data from Global Positioning Systems (GPS), it may be possible to fill this critical freight data gap. However, there is a need for methodological approaches to enable usage of this data for freight planning and operations.

To address this methodological need, we use advanced machine-learning techniques and spatial analyses to classify trucks by industry based on activity patterns derived from large streams of truck GPS data. The major components are: (1) derivation of truck activity patterns from anonymous GPS traces, (2) development of a classification model to distinguish trucks by industry, and (3) estimation of a spatio-temporal regression model to capture rerouting behavior of trucks.

First, we developed a K-means unsupervised clustering algorithm to find unique and representative daily activity patterns from GPS data. For a statewide GPS data sample, we are able to reduce over 300,000 daily patterns to a representative six patterns, thus enabling easier calibration and validation of the travel forecasting models that rely on detailed activity patterns. Next, we developed a Random Forest supervised machine learning model to classify truck daily activity patterns by industry served. The model predicts five distinct industry classes, i.e., farm products, manufacturing, chemicals, mining, and miscellaneous mixed, with 90% accuracy, filling a critical gap in our ability to tie truck movements to industry served. This ultimately allows us to build travel demand forecasting models with behavioral sensitivity. Finally, we developed a spatio-temporal model to capture truck rerouting behaviors due to weather events. The ability to model re-routing behaviors allows transportation agencies to identify operational and planning solutions that mitigate the impacts of weather on truck traffic. For freight industries, the prediction of weather impacts on truck driver’s route choices can inform a more accurate estimation of billable miles.

Share

COinS