In the present thesis we have explored the potential of Gaussian Process regression for forecasting urban traffic in a number of arterial streets in the city of Stockholm. Predictions are based on traffic speed data obtained from probes (GPS-equipped taxis) in the Södermalm district, and on weather data from a nearby meteorological station. In contrast to most of the existing literature, we propose a Bayesian non-parametric approach to predict the traffic at specific quarter-hours at any day of the week. The proposed model features an additive covariance function that encodes both short-term autocorrelations in traffic and long term effects such as the different days of the week and months by using Squared Exponential kernels with Automatic Relevance Determination.
Results shows that the proposed Gaussian Process performs as well as basic configurations of popular alternatives such as the Multilayer Perceptron or the K-Nearest Neighbor regression in terms of prediction accuracy. By interpreting the estimated kernel parameters we are able to identify weekdays or months with characteristic traffic patterns at each route, and also to shed light on the cause of apparently-odd monthly patterns in certain arterials, which were a posteriori related to roadworks by the use of traffic incident data. We also corroborated the existence of network inter-dependencies as traffic patterns in arterials nearby to roadworks are also affected. The thesis also discuss the limitations of the proposed method to address this traffic forecasting problem using real data, which are mainly related to the high computational cost for training the model.