Electricity Price Forecasting in the Swedish Day-Ahead Market : A Comparative Study of Machine Learning Models Across Multiple Forecasting Horizons

Sammanfattning

Electricity price forecasting is a challenging problem due to its high volatility and the large number of factors influencing the electricity market. This thesis investigates how different machine learning models compare in forecasting day-ahead electricity prices on the Swedish Nord Pool market and how different forecasting horizons affect model performance. Three forecasting horizons are evaluated: short-term forecasting of the next day, medium-term forecasting of the next week, and long-term forecasting of the next month. A Naïve forecaster, which predicts future prices based on previous observations, is used as a benchmark model. The evaluated methods include tree-based machine learning models, statistical models, and neural networks. The models were trained using historical electricity market data from January 2020 to March 2026 together with weather data, load forecasts, and calendar-based features such as time of day and day of week. All models were evaluated across the four Swedish bidding zones using two separate test sets, one of which was held out as a blind test until all model development had been completed. The tree-based and neural network models were also evaluated using cross-validation. Model performance was assessed using Mean Absolute Error (MAE), Normalized Mean Absolute Error (NMAE), and Root Mean Squared Error (RMSE). The results show that the neural network model LSTM achieves the strongest performance for short-term forecasting, while tree-based models perform best for the medium- and long-term forecasting horizons. The study further shows that the relative performance of forecasting models varies substantially across different forecasting horizons, highlighting the importance of selecting models based on the prediction horizon.