Travel time prediction using ensemble learning algorithms

In the research area of travel time prediction, the existing studies mainly focus on aggregated travel time prediction (without distinguishing vehicle types) or travel time prediction for passenger vehicles. The travel time prediction for freight transportation has not received enough attention from researchers. Only a few relevant studies can be found in the literature, and the proposed methods are usually very simple and lack comparisons with more advanced methods. Although many believed that a better prediction model can be developed using more advanced techniques such as artificial neural networks or autoregressive conditional heteroscedastic models, it is usually difficult and costly to train these models and the model interpretability is poor. There is a demand for `off-the-shelf' methods with good performance, ease of implementation and good model interpretability. Thus, the aims of this thesis are: (1) developing some `off-the-shelf' data-driven methods to predict travel time for freight transportation; (2) creating a comprehensive understanding of how the developed methods can be more effectively applied for general travel time prediction problems. Its two main contributions are: (1) it develops data-driven travel time prediction methods for freight transportation by utilising freight vehicles' trajectory data; (2) it investigates the relation between features and performance and discovers the combinatorial effects of features under the effects of different noise processes and different model fitting strategies. The experimental results show that useful features can be mined from the trajectory data to enhance the travel time prediction for freight transportation. The developed methods outperform some of the state-of-the-art data-driven methods