Forecasting mFRR Prices and Volumes using Machine Learning - Study of mFRR CM and EAM balancing markets in the Swedish SE3 bidding zone using LSTM and MLP neural networks

Abstract

The transition towards a more renewable and decentralized power system is increasing variability and operational complexity, leading to a growing need for balancing services. This thesis focuses on forecasting mFRR prices and volumes in the Swedish SE3 bidding zone for both the Capacity Market (CM) and Energy Activation Market (EAM). The objective is to identify relevant explanatory variables and to develop and evaluate datadriven forecasting models. Machine learning methods, including Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks, are compared against naive benchmark models. Different feature set configurations, including univariate, multivariate, and selected subsets, are evaluated using early stopping and a holdout test set. Feature selection results reveal clear differences between the CM and EAM markets. For CM, the most informative variables are autoregressive signals, demand forecasts, renewable generation forecasts, spot market indicators, and seasonal temporal patterns, suggesting that market outcomes are largely driven by relatively stable system fundamentals and expected supply-demand conditions. In contrast, EAM forecasting is dominated by short-term market signals, particularly spot prices, autoregressive signals, and dependencies between EAM prices and volumes. Wind-related variables and system-state indicators such as inertia and frequency deviation also provide additional predictive value, while temporal variables show less relevance. This reflects the more dynamic and reactive nature of the EAM market. Forecasting performance varies considerably across markets and targets, but no model architecture or feature configuration consistently outperforms the others. In the CM market, simple statistical baseline models often outperform machine learning models, suggesting that CM behavior is largely driven by recurring temporal patterns, stable system conditions, and expected future supply-demand conditions. In contrast, machine learning models outperform naive benchmarks in EAM, indicating more complex and reactive dynamics that are harder to capture with simple baselines, reflecting a distinction between longer-term planning in CM and near real-time balancing operations in EAM. Overall, mFRR forecasting performance is highly target- and market-dependent, and increasing model or feature complexity does not systematically improve predictive performance.