PREDICTING DAY AHEAD WIND PRODUCTION USING TEMPORAL FUSION TRANSFORMER

Sammanfattning

As the world is shifting away from fossil fuels and heads towards a more sustainable tomorrow, it becomes important to be able to predict the more volatile energy types such as wind and solar generation. Reliable forecasts would allow for both companies and individuals to plan their energy use daily, such as battery or EV charging. These predictions become essential as the society embraces the cleaner energy alternatives and gives people the potential to make informed decisions of their energy uses. This raises the question of how well, and how far into the future a Temporal Fusion Transformer (TFT) model can predict. The TFT is a transformer model that utilizes LSTM as encoder and decoder blocks. The experiment includes preprocessing of data which handles wind speeds, wind direction, and cyclical encoding for the timestamps. The regions the models are being tested in are Swedish Electrical Regions SE1 and SE3. The TFT performed very well in SE3, but very poorly in SE1 due to lack of weather station coverage. It was validated against the state-of-the-art RNN-LSTM which performed very well in the early hours, but poorly at the later hours. This is the only TFT research paper that handles Swedish data and shows that AI can be used to make forecasts regarding power production, and potentially prices and consumption as well. This thesis is made together with Electrification Hub.