Data-Driven Wind Turbine Power Anomaly Detection Using SCADA

Abstract

In the light of global need for renewable energy, wind energy plays a crucial role. Today, the majority of wind turbines are still being built on land. Given this critical role, accurate monitoring methods are needed to fully understand the performance within wind farms. This thesis aims to create a methodology to evaluate the power performance of wind turbines within an onshore wind farm using SCADA data. The primary objective is to identify deviations from the expected power generation patterns by using a multivariate machine learning approach. Due to the terrain complexity of onshore wind farms, a cluster-based approach is used. In total, 15 clusters consisting of 61 turbines have been evaluated. The training data is filtered by applying multiple filters with the objective of creating a normal behaviour model. The evaluation is based on a feedforward multilayer perceptron regressor to predict the power output for a test data set. A sequential methodology is explored to refine the model performance and is then applied to power performance analysis tasks to detect anomalies in the data sets. The results of the final analysis suggest that the approach taken is capable of detecting anomalies in the data. It is shown that in multiple clusters cases of under- or overperformance can be detected.