Posters

Abstract

In the context of wind resource assessment, measurement campaigns often serve as the foundation for estimating the energy yield of the planned wind farms. Once the measurement concludes, the collected wind data has to be adjusted to reflect the long-term wind climate at the site. There are many Measure-Correlate-Predict (MCP) methods to perform this task. Generally, high accuracy can be achieved when the measurement spans a full year, encompassing all relevant meteorological conditions. However, reducing the measurement duration to significantly less than one year to save time and costs introduces biases and larger uncertainties in MCP-based long-term corrections. This study presents an approach based on Similarity-Guided Ensemble Learning with Domain Generalization to mitigate the seasonal biases and uncertainties when the measurement period is limited to just a few months. The analysis utilizes wind measurement data from 41 sites across Germany. Three different reanalysis datasets serve as long-term reference data for the long-term correction.  The proposed workflow consists of four steps. First, one site and measurement period of 90 days are selected and the meteorological conditions in this period are assessed. Second, a similarity analysis is conducted, identifying comparable meteorological conditions across measurements from other locations. Third, an accuracy analysis is done, which reveals the performance of MCP-based long-term correction methods on the other sites which are included in the similarity analysis. In these, two widely used linear models, namely linear regression with residuals (LR) and variance ratio (VR) and extreme gradient boosting (XGBoost) are applied. Fourth, domain generalization with ensemble learning is performed, featuring stacking ensemble and coopetitive soft gating ensemble (CSGE). In this context, ensemble members from same model type and data but different similar periods are ensembled to perform the MCP. The study concludes that using the designed workflow can reduce both bias and systematical uncertainty coming from seasonal effects. The similarity analysis and ensemble learning contributed well to bring the most similar stations meteorologically. Also, bias results were in a very acceptable range, better than the reference results of the VR. Yet, changing the number of variables did not make a significant improvement. Overall, the encouring results from XGBoost demonstrate that Similarity-Guided Ensemble Learning with Domain Generalization can contribute significantly to a higher accuracy and lower uncertainty in wind resource assessment which is based on a measurement campaign comprising a few months only.