Presentations

Abstract

In cold climates, icing on wind turbines is a major source of production losses, second only to wake effects. Accurately accounting for these losses is crucial during both the pre-construction phase of wind farm projects to reduce uncertainty in energy yield assessments, and during operation to predict energy imbalances and minimize economic costs in energy trading.  At Kjeller Vindteknikk/Norconsult, we have developed and continuously improved a physical model, IceLoss, since 2008 to assess long-term average icing losses for wind farms in the pre-construction phase. This model has been fundamental in reducing uncertainty in energy yield assessments by providing more accurate predictions of icing-related production losses. The modelled ice mass (g/m) and icing intensity (g/m/h) on turbine blades are translated to production loss using two methods: a reduction matrix (default) and a machine-learning (ML) based solution. The reduction matrix method uses static relationships in the historical data between icing conditions and power loss, while the ML-based solution dynamically learns these relationships. By integrating ML techniques with physical models, we aim to create hybrid models that utilize the strengths of both approaches.  Building on the IceLoss model, we have developed a day-ahead icing forecast model, IceLossForecast, which gives important input for stakeholders in the power market. The performance of the IceLossForecast model has proven to be successful, with a high rate of correctly modeled icing hours and a low rate of false alarms. Ice accumulation on wind turbine blades is very sensitive to weather conditions, and traditional numerical weather prediction (NWP) models face challenges in accurately representing the vertical and horizontal location of clouds and humidity due to limited model resolution, local topography effects, and imperfect parameterizations. The forecast uncertainty is therefore assessed using an ensemble approach, which involves running multiple simulations with varying meteorological conditions and model parameters. The ensemble approach helps quantify the range of possible outcomes, and the ensemble mean provides a more robust forecast than using only the deterministic forecast. Furthermore, IceLossForecast can make use of real-time SCADA data, continuously tuning the forecasts towards the observations. The frequent updates increase the forecast accuracy and reflect the latest atmospheric conditions. By incorporating real-time SCADA data, the model can adapt to changing conditions and provide more precise predictions. The IceLossForecast model has been shown to enhance day-ahead wind energy production estimates on various scales. These improvements in forecast accuracy ultimately lead to more efficient and cost-effective energy system. An evaluation of the performance of the IceLossForecast for the whole of Finland during the winter season 2023–2024 show that the IceLossForecast decreased the mean absolute error (in MWh) by 10% overall. In our presentation, we will share our experience in running icing loss forecast services and demonstrate how ML-based modeling techniques and site-specific SCADA tuning can be integrated into an existing icing production loss modeling chain. We will also discuss the requirements for a good icing model, challenges ahead, and future development plans.