This study develops a machine-learning framework for long-term forecasting of prices

in the Nordic Frequency Containment Reserve (FCR) markets. Accurate long-term

price expectations are increasingly important for investors and system planners as the

energy transition introduces higher variability in power systems and increases the need

for flexibility resources.

The proposed approach applies a Temporal Fusion Transformer (TFT) model to

historical balancing market data together with scenario-based electricity price

trajectories from Svenska kraftnät’s Long-term Market Analysis 2021 (LMA2021).

Electricity spot price scenarios are treated as known future covariates, enabling the

model to translate projected energy market conditions into long-term forecasts of

ancillary service prices.

Historical data covering 2019–2025 were used for training and validation, including

reserve prices, reserve volumes, generation mix, hydro storage levels, and electricity

spot prices. The trained models were then applied to multiple weather-year realizations

under two system development scenarios: Electrification with Renewables (EF) and

Electrification with Plannable Generation (EP).

Results indicate that the model successfully captures broad seasonal patterns in reserve

prices but struggles to reproduce extreme price spikes due to limited explanatory

variables and short historical data series. Long-term forecasts suggest relatively stable

price ranges with moderate variability between scenarios, although results remain

highly dependent on the underlying electricity price assumptions.

The study demonstrates that scenario-driven machine-learning models can provide

useful long-term insights into balancing markets, while also highlighting the

importance of incorporating additional system drivers and longer historical datasets to

improve forecasting robustness.