Umeå University's logo

In the ongoing climate crisis, transitioning to renewable energy sources is essential to manage the increasing energy demand. One such renewable energy source is the weather-dependent energy source, wind power. Many wind farms are located in Cold Climate (CC) regions, known for their vast potential for wind power production. The weather conditions of the CC regions favor ice events, such as iceformation and accumulation on the wind turbine blades. This negatively impacts power production by changing the aerodynamic behavior and introducing an imbalance between the turbine blades. Hence, predicting ice events in advance is compelling to transition wind power into a more reliable energy source. This thesis presents a new method for predicting ice formation on wind turbine blades. It aims to improve the existing short-term wind power forecasting models by utilizing supervised Ma-chine Learning (ML). Three ML models were trained, evaluated, and compared. The three ML models examined were XGBoost (XGB) Regression, Random Forest (RF) Regression, and Support Vector Regression (SVR). The best-performing model was used for classifying ice formation. Two wind farms (wind farm A and B) of different sizes and complexities were examined.

The supervised ML models were trained by merging Supervisory Control And Data Acquisition (SCADA) data holding the power output of each wind turbine of each wind farm with local Numerical Weather Prediction (NWP) data holding atmospheric parameters of the wind farm sites. For wind farm A, it was found that the XGB Regression model was the best-performing ML model with an R2 value of 69.64% when evaluated on data ranging from August to November (non-ice data). When evaluated on data including ice events, the performance decreased, giving an R2 value of 59.56%. The best-performing model for wind farm B gave an R2 value of 38.69%, when evaluated on winter data and was not good enough to perform ice classification. The predicted values from the best-performing ML model of wind farm A were used for classifying the ice formation by estimating the percentage difference between a reference power curve and the predicted values. The predicted values falling outside the 10th percentile of the reference power curve were classified as production losses caused byice formation. Using the predicted values for the winter period, we found that ice formation could be accurately performed, where 89.33% of the predicted production losses caused by ice formation values were found during periods in which ice was detected.

From the findings, we can conclude that the investigated method of improving wind power forecasting and predicting production losses caused by ice formation on wind turbine blades shows potential. By improving the model to handle day-ahead forecasts better, the method could potentially be used to predict ice formation and accumulation on the wind turbine blades, using daily NWP data as input.