Korttidsprognos av mFRR-aktiveringsriktning och obalanspris med XGBoost : En tvåstegs flerhorisontstudie för att stödja intradagshandel på den Nordiska elmarknaden

This study investigates the use of Extreme Gradient Boosting (XGBoost), an ensemble-tree-based machine learning (ML) algorithm, for short-term forecasting of manual Frequency Restoration Reserve (mFRR) activation direction and imbalance prices in the SE2 price area of the Nordic electricity market. With growing electrification, a rising share of renewables and strain from the ongoing energy crisis, the ability to anticipate balancing needs has become both more critical and more complex. Despite its practical importance, this problem has been relatively underexplored in the academic literature. The objective of this study is to assess how ML can support Intraday (ID) decision-making and improve predictive capabilities. To achieve this, a two-stage forecasting framework was developed using XGBoost, with the first stage classifying the expected direction of mFRR activation (upregulated, downregulated or unregulated), and the second stage predicting the corresponding imbalance price. The models were trained on data collected between May 2022 and February 2025 and evaluated across multiple forecasting horizons. By systematically testing pre-processing techniques, data availability, model configurations and feature sets and validating assumptions through domain knowledge, this study makes valuable contributions to the field. The results demonstrate that creating a model with normalized transmission flows, including forecasting error features, and data from SE1, SE2, SE3 and NO4 yielded the best results. Forward filling for missing values, avoiding outlier removal, applying simple ordinal time encoding, and relying on Recursive Feature Elimination (RFE) for feature selection further improved performance. Leveraging a wide set of features and optimizing hyperparameters through Bayesian optimization, enhanced model performance. More specifically, the classification model achieved 𝐹1macro scores of 0.805, 0.611 and 0.531 for one-, two- and three-hour ahead forecasts. While predictive performance declined at longer horizons, class-specific Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) scores indicated that meaningful signals were retained up to five hours into the future and that prediction thresholds could be further optimized per class. With regards to the imbalance price model, one-hour forecasts yielded acceptable results. This model reached a Normalized Mean Absolute Error (NMAE) of 0.366 and a Coefficient of Determination (𝑅2) of 0.667 for the one-hour ahead forecast. Overall, the study concludes that XGBoost models, supported by rigorous data preprocessing, feature engineering, feature selection and hyperparameter optimization, can provide valuable short-term forecasts in the context of imbalance prediction.