Attempts at using Bayesian neural networksfor uncertainty assessments of temperature forecasts

Abstract: This thesis describes attempts at estimating the uncertainty of the 2-metre temperature forecast error from a probabilistic point of view, utilizing Bayesian neural networks. Bayesian neural networks are a type of machine-learning algorithms used to find patterns in data and make probabilistic predictions. Multiple fields of output data from the ECMWF IFS global model, along with temperature measurements from two meteorological observation stations for a period of six years are used for training of the networks. Attempts are made to assess probability distributions for the error as a continuous variable and through approaching the task as a binary classification problem. None of the attempts described were successful in terms of producing useful predictions, but may serve as a starting point for further investigations.