Forecasting Finnish aFRR energy reserve market prices using deep learning and tree-based models

With the increased integration of renewable energy sources, electricity systems face growing challenges due to the intermittency and volatility of weather-dependent generation. To maintain grid stability, reserve mechanisms such as the Automatic Frequency Restoration Reserve (aFRR) play a critical role by balancing supply and demand in real-time. However, the high volatility of aFRR energy prices complicates the decision-making of reserve providers and market operators. To address this challenge, this paper aims to forecast 48-hour aFRR marginal energy prices in the Finnish electricity market under both Down- and Up-regulations. We implement and compare two tree-based models, LightGBM and XGBoost, with two deep learning approaches, the Temporal Fusion Transformer and the Time-series Dense Encoder. The models are trained using historical Finnish aFRR energy prices, complemented by relevant market and weather variables, and evaluated across three cross-validation strategies. The models achieved mean absolute errors of 17.5–17.8 for Down-regulation and 55.7–60.1 for Up-regulation, with LightGBM delivering the most consistent performance under limited data availability. To our knowledge, this is the first study to systematically compare data-driven forecasting models for aFRR energy prices in the Finnish electricity market, providing an early-stage comparative baseline and economic relevance for a diverse array of stakeholders.