Posters

Abstract

The performance of wind energy converters (WECs) primarily depends on wind speed. Although the power curve of a WEC is determined by the manufacturer at the prototype site, the wind conditions at the actual installation site are usually different. This raises the question of whether the WEC delivers its optimal performance – and thus energy yield – under the real site conditions. For the operator, this is difficult to assess, because anemometer measurements taken directly at the turbine are always subject to uncertainties. Permanent additional measurements of the wind conditions, for example using LiDAR, incur further costs. We therefore seek a cost-efficient solution to determine the effective wind speed at the rotor using operational data from the turbine, in combination with machine learning (ML) models. This would enable WEC operators to check the performance of their turbines more cost-effectively and, if necessary, adjust turbine control to increase energy yield or to assert claims against the manufacturer.  We present a real-world case study that shows how ML-based estimation of effective wind speed provides a more accurate representation of the wind inflow conditions at the rotor compared to the turbine anemometer. This enables a more reliable assessment of turbine performance and energy yield under real site conditions. The data-driven ML method estimates effective wind speed using standard SCADA data. This data is used as model inputs, while external LiDAR measurements serve as target variable in model training. After training, the resulting approach relies exclusively on operational turbine data, avoiding the need for additional wind measurements.  Raw SCADA time series are preprocessed into a structured feature set using automated signal filtering and gap-aware identification of continuous operating periods. To assess the benefit of ML over classic approaches, three approaches of increasing complexity are evaluated: the manufacturer-provided turbine wind speed signal, a regularized multivariate linear regression, and a gradient-boosted decision tree model (XGBoost). The manufacturer wind speed signal is used as a pre-computed reference value. By contrast, the regression and ML approaches explicitly learn a relationship between available input data and wind speed. While the linear model assumes mainly simple, proportional relationships, the ML model can capture more complex effects such as non-linear behavior, interactions, and changes across operating conditions. All approaches are trained and evaluated using the same data splits over time to ensure a fair and transparent comparison.  The models are evaluated across temporal aggregations from 15s to 600s and multiple randomized train–test splits. Averaged over all aggregations and random seeds, XGBoost achieves a mean absolute error (MAE) of 0.72m/s, outperforming linear regression (0.90m/s) by 19% and the turbine sensor (0.82m/s) by 11%. The 90th percentile absolute error is reduced to 1.49m/s, compared to 1.81m/s and 1.60m/s for the linear regression and turbine sensor, respectively. At 15s resolution, MAE reductions reach 25% relative to linear regression and 15% relative to the turbine sensor. The fraction of predictions within ±1.0m/s increases to 80% for XGBoost, compared to 74% and 75%, highlighting its better performance at high temporal resolution.