Evaluating the suitability of Gaussian process regression and XGBoost on electricity price forcasting

Sammanfattning: Electricity finds itself different from other fresh-ware commodities, it cannot easily be stored. This characteristic trait of electricity results in traditional pricing methods not working for electricity pricing. Thus different pricing schemes are needed, such as Price Forward Curves (PFC) or pricing against a price level. The price forward curves are constructed through a mix of historical market data and model predictions, and the price levels are computed by dividing the price of each hour by the average monthly price to get a ratio, so called Hour-to-month ratio (H2M). This ratio can then be used instead of prices to create predictions. Furthermore, the German electricity sector is changing, with a rapid growth of renewable energy production a better understanding on how future electricity prices and how to model the future price curves is needed. In this thesis, I will first study how different energy production types work as explanatory variables through linear regression on differentiated data. That knowledge will then be taken and put it to use in Gaussian Process Regression but with H2M ratios instead of prices. Then some exploration on how to include dummy variables in Gaussian Process Regression is done, with the use of different model families to easily compare the result within each model group. Lastly a short evaluation on whether the XGBoost software is a good fit for the problem is done. This study will be done in the German power market and uses data from smard.de, which can be found in chapter 3. It shows that renewables are a good predictor and later on the discussions about the different model structures will be found.