Clustering Wind Turbines for SCADA Data-Based Fault Detection

Condition monitoring systems are commonly employed for incipient fault detection of wind turbines (WTs) to reduce downtime and increase availability. Data from supervisory control and data acquisition (SCADA) systems offer a potential monitoring solution. Multi-turbine approaches, which merge variab...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on sustainable energy 2023-01, Vol.14 (1), p.1-11
Main Authors: Du, Bojian, Narusue, Yoshiaki, Furusawa, Yoko, Nishihara, Nozomu, Indo, Kentaro, Morikawa, Hiroyuki, Iida, Makoto
Format: Article
Language:English
Subjects:
Clustering
Condition monitoring
Data acquisition
Downtime
Environmental conditions
false alarm
False alarms
Fault detection
Manual control
Neural networks
Performance degradation
SCADA data
Supervisory control and data acquisition
Terrain
Turbines
Wind farms
Wind power
Wind turbine
Wind turbines
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Condition monitoring systems are commonly employed for incipient fault detection of wind turbines (WTs) to reduce downtime and increase availability. Data from supervisory control and data acquisition (SCADA) systems offer a potential monitoring solution. Multi-turbine approaches, which merge variables recorded from different WTs on the same wind farm, have been developed to improve fault detection performance by reducing the influence of variations in environmental conditions. However, in complex terrain, environmental conditions vary among WTs. Moreover, manual control (e.g., maintenance and curtailment) can also raise false alarms in some WTs. Here, the false alarm characteristics of a wind farm in complex terrain are investigated. A clustering-based multi-turbine fault detection approach is proposed, consisting of three steps: WT clustering, single-turbine modeling, and fault indicator calculation. First, k-medoids clustering with dependent multivariate dynamic time warping is applied for WT clustering. Then, an autoregressive neural network is used to construct a single-turbine model. Finally, residuals between median values of the model output of all WTs in the same cluster and the target WT are used to calculate the anomaly level. Evaluation results for real large-scale SCADA data confirm that the proposed approach raises fewer false alarms without degrading detection performance.
ISSN:1949-3029
1949-3037
DOI:10.1109/TSTE.2022.3215672