Even though state-of-the-art weather prediction models in use today require high standards for computational power and are fed with a multitude of observation data, they are not fully capable of predicting weather for locations that fall below the resolution of their grid.
I based the work of my thesis on the hypothesis that it is possible to improve the weather prediction for the location of a given weather station compared to the output of the AROME numerical weather prediction model, solely by teaching a neural network with tuples of earlier measurements and their corresponding predictions from the past.
My goal was to create the prototype of a machine learning system that can also aid the objective measurement of the performance of its results in order to test the hypothesis.
I gave an introduction to the process of numerical weather prediction, the structure of artificial neural networks and the basics of supervised learning. Afterwards I presented the development of the prototype starting with its block diagram, through the pre-processing of data to the measurements which validated the system.
Testing the prototype with independent data has found that the system yields slight improvements over the conventional model output, thus the hypothesis has been proven for the test data set.