A substantial share of current and planned wind energy projects are located in cold climates, where ice accretion on turbine blades degrades aerodynamic performance and leads to significant power losses. Moreover, the processes of ice accretion and ablation are governed by complex, time-dependent physics, making the estimation of icing losses difficult and thereby introducing uncertainties during pre-construction energy yield assessments, as well as complicating strategic decision-making. Although several approaches currently exist, these models for estimating ice-related losses rely on input variables with high measurement uncertainty, require on-site meteorological mast data, or define icing events in ways that are inconsistent with industry standards. To address these limitations, this study proposes a machine learning framework that predicts icing losses using low uncertainty and widely available meteorological inputs - namely temperature, relative humidity, wind speed and precipitation. The framework is trained and validated using data from nine wind farms in Finland and Sweden. Two training strategies are examined: a global model that employs aggregated data from multiple sites, and a local ensemble composed of farm-specific models. Each strategy is evaluated using three learning architectures: linear regression, Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosted Trees (XGBoost). Among all combinations, the global XGBoost model demonstrates the strongest performance, achieving a relative mean absolute error of 22.67% in total energy loss estimation across the dataset, closely matching the capabilities of commercial state-of-the-art tools. Interpretability analysis further indicates that the model generally relies on features in a manner consistent with the established physical understanding, thereby enhancing confidence in its predictive ability. Taken together, the proposed framework offers a robust and practical approach for quantifying icing losses during pre-construction energy yield assessments, providing a valuable tool for wind energy developers in cold climates. • A machine learning framework is developed to predict wind turbine icing losses using simple, low-uncertainty mesoscale meteorological inputs. • Two training strategies and three model architectures are evaluated using data from nine wind farms in Sweden and Finland. • The best model achieves a 22.67% relative mean absolute error in total icing-loss estimation, matching the performance of commercial state-of-the-art tools. • Feature analysis shows strong alignment between model behaviour and established icing physics.
Stefanov et al. (Wed,) studied this question.