Presentations

Abstract

In the era of digitalisation, an abundance of data is available that data-driven approaches, especially deep learning methods, can leverage. In the recent years those approaches have been employed more and more in the wind energy sector. The reliability of these methods, however, depends on high-quality data. In the case of wind turbine generators (WTGs), data comes from the Supervisory Control and Data Acquisition (SCADA) system, containing variables such as wind speed, wind direction, power output and more. However, faulty and missing data arise due to sensor errors, software issues, icing, dirt, or even human intervention, which reduces data quality and hence impacts data-driven model performance. Missing data can be dealt with by either deletion or imputation, with the former discarding all data where a missing value is observed and the latter trying to find appropriate values to fill those gaps. Recent advances in data imputation strategies include the use of Graph Neural Networks (GNNs). The advantage of GNNs compared to other common machine learning architectures such as Support Vector Machines, Random Forest Trees and more is that local and global spatial relationships between WTGs can be easily captured. In recent years they have shown to taking into account and predicting the influencing patterns between turbines. Within this work we employ GNNs for imputing missing wind speed data based on the open-source dataset of the Kelmarsh wind farm in the UK. We talk about the model architecture and the model training as well as analyse the results and interpret the specific model outputs and information.