Machine learning methods for seasonal forecasts of climate variables

The purpose of this thesis work is to give an overview of the most prominent and successful machine learning models in the field of climate forecasts, to exhibit some of the most important methods used in their implementation, and to showcase their performance through the use of a few simple example models. We first go through the inner workings of the LSTM and transformer models, highlighting their strengths and shortcomings. We then go on to the data preprocessing phase, which in our case included the EOF decomposition of our input fields, particularly the tropical sea surface temperature and surface air temperature. We also list some practical methods that are useful during the training process. Finally, we present the performance of our example LSTM and transformer models.