Spatio-Temporal Prediction for Mobile Network Quality Analysis

Abstract

Mobile networks are under constant change and require frequent optimization and replanning. Towards this goal there are multiple tools however, the data coming from the user equipment is underutilized for this purpose. In this thesis the measurements collected from user devices are exploited to create spatio-temporal maps, following ideas coming from radio environment maps literature, that can be used for optimization and planning. The thesis studies different ways of creating such maps and predictions and extends them to forecasting in the temporal domain. The prediction results are visualized so that network operators can use them for optimization and planning. Locally approximate Gaussian process regression is a novel approach in this context and it is studied along with k-nearest neighbour interpolation, fixed rank kriging and a neural network based solution. The novel approach is among the best performing overall.