Data imputation for SCADA data using Graph Neural Networks

Abstract

Abstract Missing data in wind turbine SCADA systems can arise due to sensor failures, software issues, or maintenance, impacting data mining and analysis tasks. This work investigates the use of Graph Neural Networks (GNNs) for imputing missing wind speed data by leveraging global and local spatial relationships between turbines. The goal is to improve data quality and completeness for downstream tasks such as energy loss estimations and performance analysis. A GNN model was developed using SCADA data from the Kelmarsh wind farm, incorporating wind speed components and turbine nacelle orientations as node features. The model was trained to predict missing wind speeds using information from neighbouring turbines. Its performance was compared against three baseline methods, namely a simple mean model, a k-nearest neighbors (KNN) imputer, and a Generative Adversarial Imputation Network (GAIN). Results show that the GNN model improves wind speed prediction accuracy, measured by the root mean square error, by around 20% and downstream power predictions, evaluated using the mean absolute percentage error, by around 3% compared to the GAIN, which was the second best model. However, we also found that more accurate wind speed and power predictions do not necessarily result in more accurate energy loss estimations and Weibull distribution fitting, likely due to over-estimations and under-estimations cancelling each other out.