Posters

Abstract

Keywords: Site assessment, machine learning, explainable AI, feature importance, partial dependence plot Abstract As the wind energy industry undergoes digital transformation, emerging technologies like artificial intelligence (AI) have become crucial for optimizing various stages of a wind energy project. This research focuses on applying machine learning (ML), specifically explainable AI (XAI) techniques, to advance the mesoscale site assessment process and make it more resource-efficient. Material and Methods The research is based on a dataset of nine input variables, including turbine-specific features (e.g. rotor diameter) and location-specific features (e.g. wind speed) of 5,574 wind turbines in Denmark. The capacity factor of the year 2019 is taken as the outcome/dependent variable. This dataset served for an a posteriori analysis, where one Random Forest Regressor, one Gradient Boosting Regressor, and two different Extreme Gradient Boosting Regressors (XGBR1 & 2) were trained. On top of these regressors, the two XAI methods, Permutation Feature Importance (PFI) and Partial Dependence Plots (PDP), are applied. By measuring the increase in the prediction error of a model when breaking the relationship between one feature and the corresponding outcome, PFI identifies the features in the dataset that are most important for the model when making predictions. In this case, this will show which features have the highest impact on the capacity factor of the wind turbines. Moreover, PDPs illustrate the marginal effect of selected features on the predicted outcome in an all-else-equal approach (ceteris paribus). This means that in this way, it can be identified how the capacity factor of a turbine would change if its latitude/longitude changes. Results, Discussion and Contribution All the trained models have an R2 score greater than 84%. Based on the best-performing model, the XGBR2, the most important feature is the rotor diameter, followed by the capacity, and the wind speed is among the least important ones. These results make great sense for a country like Denmark, which has high and stable wind speeds nearly everywhere, and the terrains are optimal for wind turbines. Furthermore, the PDP identifies the northwestern part of Denmark as the most optimal location to place the turbines, confirming existing knowledge but providing a more resource-efficient approach to reach these conclusions. On the one hand, based on existing data, this study gives insight into the features that significantly influence the capacity factor of wind turbines. On the other hand, it sheds light on the most optimal geographical-mesoscale locations for wind turbine placement . This validates the reliability of ML for mesoscale site assessment and provides a valuable methodology to gain insights into regions newer to the wind energy industry in a resource-efficient way. Because of its transferability, this study offers valuable knowledge for global wind energy expansion.