UVM Theses and Dissertations
Format:
Print
Author:
Mark, Charles D.
Dept./Program:
Civil and Environmental Engineering
Year:
2005
Degree:
MS
Abstract:
The implementation of Intelligent Transportation Systems (ITS) in recent years has resulted in the development of systems capable of monitoring roadway conditions and disseminating traffic information to travelers in a network. However, the development of algorithms and methodologies specialized in handling large amounts of data for the purpose of real-time control has lagged behind the sensing and communication technological developments in ITS. This thesis focuses on the development of artificial neural network (ANN) algorithms capable of predicting experienced travel time during non-recurrent congestion, such as capacity reducing traffic accidents, or use in assisting Dynamic Route Guidance Systems.
Four conference papers and journal articles comprise the main body of the text. The first article is a preliminary study, which uses data generated by the Cell Transmission Model (a macroscopic traffic model) to explore the capabilities and influences of various factors when ANNs are implemented in predicting travel time. The next journal article's goal is to refine the techniques and methodologies developed in the first study. PARAMICS, an advanced microscopic traffic model, is implemented to produce data from a model of a segment of I-89 near Burlington, Vermont. The third journal article incorporates ANNs to assist in calibrating the Cell Transmission Model to the PARAMICS model: The fourth journal article extends the freeway travel time prediction techniques and methodologies developed in the first two studies to a signalized arterial corridor connecting Winooski and Essex Junction, Vermont.
Foremost, the computational experiments in the four journal articles demonstrate that ANNs are capable of predicting experienced travel time using real-time data from loop detectors embedded in the transportation network. The ANNs studied were capable of reliably predicting experienced travel time within approximately 4% of the actual travel time for freeway transportation networks and within 9% of the actual travel time for signalized arterial corridors. The ability of ANNs to predict experienced travel time during non-recurrent congestion conditions is enhanced by including a greater proportion of time periods in the training set corresponding to congested conditions compared to uncongested time periods.
Computational experiments prove that speed data is the most important and influential data type when real-time loop detector data is used by ANNs to predict experienced travel time. Additionally, choosing an ANN topology that is directly able to incorporate previous temporal states of the system is unnecessary, since the ANN is, able to indirectly account for temporal effects through the spatial dimension of the data. ANNs also proved a useful non-linear regression tool in the process of calibrating the Cell Transmission Model to the PARAMICS model. Differences between the two models were reduced from 9.2% after an exhaustive search calibration to 5.1 % after an ANN was used to adjust the output from the Cell Transmission Model.
Four conference papers and journal articles comprise the main body of the text. The first article is a preliminary study, which uses data generated by the Cell Transmission Model (a macroscopic traffic model) to explore the capabilities and influences of various factors when ANNs are implemented in predicting travel time. The next journal article's goal is to refine the techniques and methodologies developed in the first study. PARAMICS, an advanced microscopic traffic model, is implemented to produce data from a model of a segment of I-89 near Burlington, Vermont. The third journal article incorporates ANNs to assist in calibrating the Cell Transmission Model to the PARAMICS model: The fourth journal article extends the freeway travel time prediction techniques and methodologies developed in the first two studies to a signalized arterial corridor connecting Winooski and Essex Junction, Vermont.
Foremost, the computational experiments in the four journal articles demonstrate that ANNs are capable of predicting experienced travel time using real-time data from loop detectors embedded in the transportation network. The ANNs studied were capable of reliably predicting experienced travel time within approximately 4% of the actual travel time for freeway transportation networks and within 9% of the actual travel time for signalized arterial corridors. The ability of ANNs to predict experienced travel time during non-recurrent congestion conditions is enhanced by including a greater proportion of time periods in the training set corresponding to congested conditions compared to uncongested time periods.
Computational experiments prove that speed data is the most important and influential data type when real-time loop detector data is used by ANNs to predict experienced travel time. Additionally, choosing an ANN topology that is directly able to incorporate previous temporal states of the system is unnecessary, since the ANN is, able to indirectly account for temporal effects through the spatial dimension of the data. ANNs also proved a useful non-linear regression tool in the process of calibrating the Cell Transmission Model to the PARAMICS model. Differences between the two models were reduced from 9.2% after an exhaustive search calibration to 5.1 % after an ANN was used to adjust the output from the Cell Transmission Model.