Shape based classification and functional forecast of traffic flow profiles

This dissertation proposes a methodology for traffic flow pattern analysis, its validation, and forecasting. The shape of the daily traffic flows are directly related to the commuter's traffic behavior which merit analysis based on their shape characteristics. As a departure from the traditional approaches, this research proposed a methodology based on shape for traffic flow analysis. Specifically, Granulometric Size Distributions (GSDs) were used to achieve classification of daily traffic flow patterns. A mathematical morphology method was used that allows the clustering of shapes. The proposed methodology leads to discovery of interesting daily traffic phenomena such as five normal daily traffic shapes beside abnormal shapes representing accidents, congestion behavior, peak time fluctuations, and malfunctioning sensors.;To ascertain the significance of shape in traffic analysis, the proposed methodology was validated through a comparative classification analysis of the original data and GSD transformed data using the Back Prorogation Neural Network (BPNN). Results demonstrated that through shape based clustering more appropriate grouping can be accomplished that can result in better estimates of model parameters.;Lastly, a functional time series approach was proposed to forecast traffic flow for short and medium-term horizons. It is based on functional principal components decomposition to forecast three different traffic scenarios. Real-time forecast scenarios of partially observed traffic profiles through Penalized Least squares (PLS) technique were also demonstrated. Functional methods outperform the conventional ARIMA model in both short and medium-term forecast horizons. In addition, performance of functional methods in forecasting beyond one hour was also found to be robust and consistent.