Posters

Abstract

Keywords: Site assessment, machine learning, explainable AI, SHAP, Counterfactual explanation Abstract Even though a vast amount of (operational) data is generated by wind turbines, and the wind energy industry has digitalized a lot using emerging technologies like artificial intelligence (AI) and the Internet of Things (IoT), transforming these data into actionable insights still remains a remarkable challenge. Given this, the following research utilizes explainable AI (XAI) techniques and more concretely local explanations to provide valuable insight into the parameters that contribute to the performance of each turbine and deliver actionable recommendations for performance optimization. The aim of the research is not only to optimize the operational performance of existing turbines but also to enhance the site assessment process for future turbines on a micro-scale. Moreover, this research highlights the value of integrating interpretability into wind turbine data analytics to enable better-informed decision-making. Material and Methods The dataset used to train the machine learning (ML) models contains turbine-specific and location-specific parameters of more than 5000 turbines in Denmark, and the performance of each turbine in the form of the capacity factor is taken as the output variable. Using this dataset, three different ML models, namely Random Forest Regressor, Gradient Boosting Regressor, and Extreme Gradient Boosting Regressor, were trained. These models were trained to predict the capacity factor of each turbine based on the other, provided input parameters like rotor diameter, wind speed, etc. On top of that, Shapley additive explanations (SHAP) were used to examine how each parameter influences a given turbine's performance on a local level. Counterfactual explanations, on the other hand, built upon these findings and predicted specific adjustments that would increase the capacity factor considering the real-world constraints.   Results, Discussion and Contribution All the trained ML models have an R2 score higher than 80%, making the findings valid. The results from the SHAP analysis showcase to what extent the rotor diameter and/or wind speed have a positive, respectively, negative impact on the performance of a selected individual turbine, which helped to identify the reasons behind the performance of the best-performing turbines. Furthermore, the counterfactual explanation highlights how much, e.g. the rotor diameter of the least performing turbines needs to be increased in order to get a capacity factor that would be, e.g. 10/20 per cent higher than the capacity factor it has now. For the time being, these results will help better understand the operating turbines' performance and how these could be improved. However, in the future, the same model and the XAI techniques could be used to identify the optimal turbine setting, e.g., rotor diameter for new turbine locations or repowering.