Multi-fidelity spatial regression for air temperature predictions using first, second and third party data

Abstract

There has been a growing interest in using second and third party data in addition to first party data to make weather predictions. This development is necessary, because it can provide predictions for a higher spatial resolution than the predictions made using only first party data. Consequently, weather phenomena with a high spatial variability such as rain and wind could be predicted more accurately. However, it has not been shown before that using WOW-data as third party data leads to accurate predictions. In this thesis a modified version of the interpolation method of Kriging will be used to demonstrate the value of WOW-data in weather predictions. The Kriging procedure is modified such that it will be able to work with noisy data and such that it can differentiate between the systematic and random errors in first, second and third party data. The robustness of the method is examined by using synthetic data such that the model can be tested for weather of various spatial variabilities. These tests show that up to a certain spatial variability first, first and second, and first, second and third party data perform equally well. However, after a certain threshold is reached ,first and second, and first, second and third party data perform better than just first party data.