Analyzing, Forecasting, and Explaining One-Day-Ahead Electricity Prices in the Swedish Bidding Zone SE3 Using Black-Box Models

Sammanfattning

Electricity Price Forecasting (EPF) is vital for modern society. While EPF literature has traditionally focused on the accuracy of the proposed models, considerably less attention has been given to explaining the results. Recent advances in explainable AI (xAI) enable deeper interpretations of complex models' behavior. However, few research papers use xAI in the context of EPF. This thesis aims to address not only gaps in the existing EPF literature but also the practical needs of market participants and system operators: by using xAI as a tool to increase the transparency, trustworthiness, and accountability of EPF made using black-box models, and by identifying driving factors that are often neglected in conventional studies. A dataset with diverse variables across multiple interconnected bidding zones was first collected and carefully prepared. Hourly data for all Nordic bidding zones (except Iceland) were gathered for a broad range of variables, including price, load, generation, and temperature, for the period from 2022-01-01 to 2025-04-15. All data, except for the SE3 price, along with a few engineered features, were used as exogenous variables in forecasting day-ahead electricity prices for the Swedish bidding zone SE3. Three black-box models, XGBoost, LSTM, and Transformer, were applied. Once satisfactory model performance was achieved, Permutation Feature Importance (PFI) was estimated for all explanatory variables to understand and interpret the models' results. The PFI results show that among the studied variables, the exogenous information that matters most for SE3 day-ahead prices is calendar seasonality (hour and weekday), complemented by a local SE3 load and a limited set of Nordic cross-border demand, generation, and weather signals. Moreover, the PFI results show that the relative importance of the exogenous variables is model-dependent; some variables are crucial for one architecture but much less important for another. The thesis results imply that in order to optimize model performance: (1) the data choice must be tailored to the modeling approach, (2) careful feature design and pruning remain important, and (3) market integration should be considered when forecasting electricity prices. Overall, the thesis demonstrates how xAI techniques can illuminate black-box models and thus provide actionable insights for practitioners.