Machine Learning-Based Prediction and Key Drivers of mFRR Activations : A Swedish Balancing Market Study

The Swedish electricity market faces increasing complexity driven by variable renewable energy integration. As the system transitions toward carbon neutrality by 2045, accurate forecasting of manual Frequency Restoration Reserve (mFRR) activation volumes has become important for maintaining system stability and efficient market operation. This thesis investigates the predictive capability of machine learning models for mFRR activation volumes across Swedish bidding zones and identifies the key system parameters driving these activations. The research employs Long Short-Term Memory (LSTM) neural networks combined with interpretability analysis to forecast 1-hour-ahead mFRR activation volumes using publicly available ENTSO-E data spanning 2014–2025. The methodology integrates feature attribution techniques including Integrated Gradients, SHAP, and Feature Ablation to identify influential system parameters. The analysis covers all four Swedish bidding zones (SE1-SE4) for both upward and downward regulation, with case studies examining extreme activation events of downward regulation in bidding zone SE2 from September 2023 and April 2024. The results demonstrate that LSTM models can effectively forecast mFRR volumes, but performance varies dramatically across regions and directions. Price region SE2 exhibits superior forecasting accuracy for downward regulation, consistently outperforming naive baseline models, while southern zones SE3 and SE4 show limited predictability due to high zero-activation ratios. The geographical asymmetry stems from concentrated hydropower resources in northern zones, which supply the majority of balancing energy. The interpretability analysis reveals consistent key drivers across extreme periods: physical flows, scheduled generation, net export positions, and actual system load dominate during stress events. Case studies confirm that model priorities align with independent market reports, demonstrating that the LSTM learns genuine system relationships. The findings establish that data-driven approaches can reliably identify physically meaningful system drivers during critical periods. The research demonstrates that zone-specific modeling strategies are necessary for heterogeneous energy systems. For system operators, the results provide actionable insights for monitoring high-impact variables, as errors in physical flows, scheduled generation, and load forecasts propagate disproportionately into activation needs