Module Description
The growth of wind energy systems, especially offshore, demands advanced
monitoring and maintenance strategies. SCADA systems are crucial for monitoring
wind turbine health, generating extensive sensor data. Anomaly detection
techniques applied to this data can proactively identify deviations from normal
behaviour, signalling potential failures. Traditional model-based fault
detection methods often struggle with the complexity of wind turbine systems.
Consequently, there's increasing interest in data-driven approaches,
particularly machine learning and deep learning, for anomaly detection in wind
energy applications. An unsupervised, multivariate anomaly detection model for
horizontal axis wind turbines using an autoencoder architecture is developed.
The research analyses SCADA data from offshore wind turbines in the Gulf of
Guinea, provided by EDP Renewables. The dataset undergoes preprocessing, and
the effect of multiple feature engineering is evaluated. Additionally, the
impact of additional anomaly labelling is evaluated by including an isolation
forest algorithm and the warning log information. The autoencoder effectively
identifies anomalies correlated with logged warnings and failure events. Model
performance is evaluated using standard machine learning metrics. Finally,
CUSUM and EWMA control charts provide deeper insights into anomaly behaviour.
Learning Outcomes
This module will introduce data driven fault detection
methods, provide a deep insight on autoencoder
architecture and deployment for anomaly detection, and compare model performance and estimate the prediction quality. Learners will be able to:
- Understand how to setup, evaluate and compare an autoencoder architecture for fault detection.
- Define a normal behaviour model
for wind turbines from data.
