Long-term correction of short-term wind measurements using Similarity-Guided Multi-Source Ensemble Learning with Domain Generalization

There are many Measure-Correlate-Predict (MCP) methods produced to make corrections to the long-term condition of a site where a wind project is planned. In this thesis project, a workflow of Similarity-Guided Ensemble Learning with Domain Generalization is introduced to have long-term correction with less bias and seasonality when the measurement period is only a few months. In this framework, the workflow consists of three steps: similarity analysis by meteorological conditions of the site, accuracy analysis, and domain generalization with ensemble learning. In these, two widely used linear models, linear regression with residuals (LR) and variance ratio (VR) for MCP, and extreme gradient boosting (XGBoost) are used as models, and stacking ensemble and coopetitive soft gating ensemble (CSGE) are experimented with for ensemble learning. Wind measurement data from 41 sites across Germany and, as reference (long-term) wind data, three different types of different reanalysis and model output data are used. 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, the promising results of the XGBoost showed that 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.