This is the author accepted manuscript. The final version is available from ASME via the DOI in this record
Offshore wind assets have reached multi-GW scale and additional
capacity is being installed and developed. To achieve demanding
cost of energy targets, awarded by competitive auctions,
the operation and maintenance (O&M) of these assets has to become
increasingly efficient, whilst ensuring compliance and effectiveness.
Existing offshore wind farm assets generate a significant
amount of inhomogeneous data related to O&M processes.
These data contain rich information about the condition of the
assets, which is rarely fully utilized by the operators and service
providers. Academic and industrial research and development
efforts have led to a suite of tools trying to apply sensor data
and build machine learning models to diagnose, trend and predict
component failures. This study presents a decision support
framework incorporating a range of different supervised and unsupervised
learning algorithms. The aim is to provide guidance
for asset owners on how to select the most relevant datasets, apply
and choose the different machine learning algorithms and
how to integrate the data stream with daily maintenance procedures.
The presented methodology is tested on a real case example
of an offshore wind turbine gearbox replacement at Teesside
offshore wind farm. The study uses kNN and SVM algorithms to
detect the fault using SCADA data and an autoregressive model
for the CMS data. The implementation of all the algorithms has
resulting in an accuracy higher than 94%. The results of this
paper will be of interest to offshore wind farm developers and operators to streamline and optimize their O&M planning activities
for their assets and reduce the associated costs.
This research was funded by the Energy Technology Institute
and the RCUK Energy Programme (Grant number:
EP/J500847/1) and EDF Energy