Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.781643
Title: Gaussian process models for SCADA data based wind turbine performance/condition monitoring
Author: Pandit, Ravi Kumar
ISNI:       0000 0004 7967 2650
Awarding Body: University of Strathclyde
Current Institution: University of Strathclyde
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
Wind energy has seen remarkable growth in the past decade, and installed wind turbine capacity is increasing significantly every year around the globe. The presence of an excellent offshore wind resource and the need to reduce carbon emissions from electricity generation are driving policy to increase offshore wind generation capacity in UK waters. Logistic and transport issues make offshore maintenance costlier than onshore and availability correspondingly lower, and as a result, there is a growing interest in wind turbine condition monitoring allowing condition based, rather than corrective or scheduled, maintenance. Offshore wind turbine manufacturers are constantly increasing the rated size the turbines, and also their hub height in order to access higher wind speeds with lower turbulence. However, such scaling up leads to significant increments in terms of materials for both tower structure and foundations, and also costs required for transportation, installation, and maintenance. Wind turbines are costly affairs that comprise several complex systems connected altogether (e.g., hub, drive shaft, gearbox, generator, yaw system, electric drive and so on). The unexpected failure of these components can cause significant machine unavailability and/or damage to other components. This ultimately increases the operation and maintenance (O&M) cost and subsequently cost of energy (COE). Therefore, identifying faults at an early stage before catastrophic damage occurs is the primary objective associated with wind turbine condition monitoring. Existing wind turbine condition monitoring strategies, for example, vibration signal analysis and oil debris detection, require costly sensors. The additional costs can be significant depending upon the number of wind turbines typically deployed in offshore wind farms and also, costly expertise is generally required to interpret the results. By contrast, Supervisory Control and Data Acquisition (SCADA) data analysis based condition monitoring could underpin condition based maintenance with little or no additional cost to the wind farm operator. A Gaussian process (GP) is a stochastic, nonlinear and nonparametric model whose distribution function is the joint distribution of a collection of random variables; it is widely suitable for classification and regression problems. GP is a machine learning algorithm that uses a measure of similarity between subsequent data points (via covariance functions) to fit and or estimate the future value from a training dataset. GP models have been applied to numerous multivariate and multi-task problems including spatial and spatiotemporal contexts. Furthermore, GP models have been applied to electricity price and residential probabilistic load forecasting, solar power forecasting. However, the application of GPs to wind turbine condition monitoring has to date been limited and not much explored. This thesis focuses on GP based wind turbine condition monitoring that utilises data from SCADA systems exclusively. The selection of the covariance function greatly influences GP model accuracy. A comparative analysis of different covariance functions for GP models is presented with an in-depth analysis of popularly used stationary covariance functions. Based on this analysis, a suitable covariance function is selected for constructing a GP model-based fault detection algorithm for wind turbine condition monitoring. By comparing incoming operational SCADA data, effective component condition indicators can be derived where the reference model is based on SCADA data from a healthy turbine constructed and compared against incoming data from a faulty turbine. In this thesis, a GP algorithm is constructed with suitable covariance function to detect incipient turbine operational faults or failures before they result in catastrophic damage so that preventative maintenance can be scheduled in a timely manner. In order to judge GP model effectiveness, two other methods, based on binning, have been tested and compared with the GP based algorithm. This thesis also considers a range of critical turbine parameters and their impact on the GP fault detection algorithm. Power is well known to be influenced by air density, and this is reflected in the IEC Standard air density correction procedure. Hence, the proper selection of an air density correction approach can improve the power curve model. This thesis addresses this, explores the different types of air density correction approach, and suggests the best way to incorporate these in the GP models to improve accuracy and reduce uncertainty. Finally, a SCADA data based fault detection algorithm is constructed to detect failures caused by the yaw misalignment. Two fault detection algorithms based on IEC binning methods (widely used within the wind industry) are developed to assess the performance of the GP based fault detection algorithm in terms of their capability to detect in advance (and by how much) signs of failure, and also their false positive rate by making use of extensive SCADA data and turbine fault and repair logs. GP models are robust in identifying early anomalies/failures that cause the wind turbine to underperform. This early detection is helpful in preventing machines to reach the catastrophic stage and allow enough time to undertake scheduled maintenance, which ultimately reduces the O&M, cost and maximises the power performance of wind turbines. Overall, results demonstrate the effectiveness of the GP algorithm in improving the performance of wind turbines through condition monitoring.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.781643  DOI: Not available
Share: