Exploring Machine Learning for Electricity Price Forecasting

Abstract

The aim of this thesis is to explore the capabilities of Machine Learning algorithms in the task of electricity price forecasting. The focus is on the Hungarian wholesale electricity market (HUPX), which is considered a benchmark power exchange in the region of SE Europe. Taking advantage of the available domain expertise, a really extended dataset was built, consisting of 69 features. For the scope of this paper, several traditional machine learning algorithms as well as artifi cial neural networks were implemented, using some well-known python libraries such as scikit-learn and keras. Moving from traditional to more sophisticated methods, it turns out that performance is constantly improving. Starting with a MAPE of 15% we managed to get down to the levels of 6% MAPE, thanks to the contribution of arti ficial neural networks, which proved their capabilities to effectively approximate a mapping function from input variables to output variable. In our effort to quantify the impact of domain expertise on the shaping of the results, a sensitivity analysis was performed, which con firmed the signifi cant contribution of each feature category to improving the performance of the algorithms. Finally, taking into account the results of other price forecasting studies in the Balkan markets, HUPX is concluded to be the most predictable power exchange, which is probably explained by the greater maturity of this power market.