Master Thesis: Icing Power Loss Forecasting for Wind Farms in Cold Climates Using Machine Learning

Published:

Master Thesis, DENSYS Erasmus Mundus Joint Master Degree Presented to the DENSYS committee in Barcelona, July 2026.

Carried out at rebase.energy, Stockholm, Sweden, under industrial supervision from Sebastian Haglund (CEO & Co-Founder) and Ilias Dimoulkas (Data Scientist), and academic supervision from Dr. Giuseppe Giorgi at Politecnico di Torino.

Abstract

Atmospheric turbine icing in cold climates causes aerodynamic degradation, structural standstills, energy losses, and operational uncertainty. This thesis forecasts wind power icing loss at 1- to 36-hour lead times using multi-year Swedish turbine SCADA and ERA5-Land data. Icing labels are established using an extended IEA Wind Task 19 Sigmoid Performance Ratio framework. Data investigation reveals inter-annual non-stationarity and wind speed–temperature overlap between icing and non-icing events. Biases between SCADA and ERA5-Land inputs are corrected using Quantile Mapping and LightGBM. From numerous candidate features across multiple physical groups, Spearman and SHAP screening remove unstable predictors. A two-stage LightGBM classifier-regressor is trained using walk-forward cross-validation over multiple winters.

Evaluation on a blind winter test set spans several scenarios: Persistence, Oracle, Realistic NWP, and SCADA-only. Persistence yields high short-term accuracy but collapses rapidly as lead time increases. Oracle (using weather observations) defines the performance ceiling. Realistic NWP tracks the Oracle classifier but suffers a regression penalty from forecast errors. SCADA-only (no weather forecasts) collapses beyond the near term. Thus, local SCADA signals dominate near-term forecasts, but meteorological forecasts are essential for longer horizons.

What I Did

  • Icing Detection & Feature Engineering: Established icing labels on multi-year Swedish turbine SCADA data using an extended IEA Wind Task 19 Sigmoid Performance Ratio framework. Diagnosed inter-annual non-stationarity and wind speed-temperature overlap between icing and non-icing regimes, and corrected SCADA-ERA5-Land biases via Quantile Mapping and LightGBM. Applied Spearman correlation and SHAP screening across multiple physical feature groups to remove unstable predictors.
  • Icing Power-Loss Forecasting: Built a two-stage LightGBM classifier-regressor, trained with walk-forward cross-validation over multiple winters, to forecast wind power icing losses at 1-36 hour lead times.
  • Scenario Benchmarking: Compared Persistence, Oracle, Realistic NWP, and SCADA-only scenarios on a blind winter test set, showing local SCADA signals dominate near-term skill while NWP forecasts are essential for longer horizons — a key trade-off for operational resilience under changing conditions.

Documentation

Next Steps

This thesis is planned to be developed into a publication — stay tuned.

Acknowledgements

Thank you to my two industrial supervisors at rebase.energy, Sebastian Haglund (CEO & Co-Founder) and Ilias Dimoulkas (Data Scientist), for their invaluable guidance and continuous feedback, and to all my colleagues at rebase.energy for a great working environment and everything I learned along the way.

I am also grateful to my academic supervisor, Dr. Giuseppe Giorgi at Politecnico di Torino, for his comments and support throughout my thesis.

To the DENSYS consortium — Fabrice Lemoine, Heathcliff Demaie, Marta Gandiglio, Samira Daghmous Menouar — and all the friends I’ve met through two years together, thank you.