Data-Driven Travel Time Prediction

Abstract: Being able to accurately forecast the time of arrival of a vehicle in traffic appeals both to private drivers aiming to keep up with their schedules, and to businesses that need to organize transport logistics. THis thesis is assigned by the Swedish truck manufacturer Scania CV AB, and sets out to use GPS data from Scania's vehicle fleet to train Machine LEarning models to predict the travel times of vehicles between stops. The predictive models implemented train on features engineered from quite simple information from the vehicles, yet reach high predictive accuracy in certain scenarios. Two approaches to predicting travel time are tested, one referred to as the Local Models approach, and the other as the Global Model approach. In the Local Models approach, many separate regressors are trained on geographical subsets of the datra and then comnbined to give global predictions. In the Global Model approach, a single regressor trains on the entire data set. The Global Model approch gives better performance than that of the Local Models in the experiment, but the Local Models approach shows some promising tendencies. It is found that a regressor predicts significantly more accurately when the geographical spread of the data is limited.