Abstract

As the share of Variable Renewable Energy (VRE) sources, such as wind, expands globally, its fluctuating nature introduces inherent volatility into power systems. In Sweden, wind production for 2025 reached 39 TWh, accounting for 24.3% of total national generation. Low VRE marginal costs depress electricity prices during high production, while transmission constraints across Sweden’s four bidding areas cause regional price divergences. In response, operators may curtail production during non-profitable periods. Quantifying this unrealized potential is vital for understanding grid inefficiencies and optimizing system operations to reach zero-emission climate goals. This study proposes a novel machine learning framework designed to reconstruct the historical technical potential of wind power generation independent of electricity market dynamics.

Applied to the aggregated production data for Sweden’s bidding areas, the methodology utilizes region-specific Histogram-based Gradient Boosting Regression (HGBR) models tuned with the Bayesian optimization framework Optuna, with meteorological features dimensionally reduced using Principal Component Analysis. The machine learning models were trained exclusively on data from profitable hours (>3 €/MWh) to prevent the algorithms from learning price driven curtailment, isolating the relationships between meteorological variables, technical availability, and production. The models were then deployed to predict the unconstrained technical production capacity across all hours, including previously unseen non-profitable hours. Subsequently, these predictions were refined using a three constraint transformation approach combining model outputs with actual production data to estimate the true technical potential.

To validate the price independence of the reconstructed technical potential, secondary validation models were trained on the reconstructed datasets and evaluated against baseline configurations, across feature sets that either included or omitted price variables. The evaluation used MAE, RMSE, and R2 metrics to verify whether price data was required to maintain predictive accuracy across both model groups. Furthermore, the SHapley Additive eXplanations (SHAP) framework was leveraged to interpret model behaviour and how price dynamics impacted performance.

The empirical results reveal that an estimated 2.90 TWh of technically feasible energy was curtailed in Sweden during 2025. Notably, 87.13% of this volume (2.53 TWh) occurred when Day-Ahead prices fell below 3 €/MWh, confirming that market price highly influence wind asset operations and highlighting future integration challenges. Regional imbalances were apparent, SE2 alone accounted for 64.77% of total curtailment, followed by SE1 at 23.16%, while the southern areas exhibited lower curtailment volumes. This extensive deviation underscores the need for targeted grid investment in transmission capacity and smart energy storage solutions