A Review and Methodology Development for Remaining Useful Life Prediction of

Offshore Fixed and Floating Wind turbine Power Converter with Digital Twin
Technology Perspective

Krishnamoorthi Sivalingam Mark Spring

Lloyds Register Global Technology Center Pvt Ltd Low Carbon Power Generation, Renewables O&M
1 Fusionopolis Place, #09-11 Galaxis, Lloyd’s Register EMEA 27 Gay Street,
Singapore- 138522 Bath, BA1 2PD, U.K.
e-mail: [email protected], United Kingdom
[email protected] e-mail: [email protected]

Marco Sepulveda Peter Davies

Industrial Doctorate Centre for Offshore Renewable Lloyds Register EMEA, Aberdeen AB10 1NN
Energy (IDCORE) - (The University of Edinburgh, AB10 1NN
University of Strathclyde, University of Exeter) United Kingdom
United Kingdom e-mail: [email protected]
e-mail: [email protected]

Abstract—The growing number of Offshore wind farms x Chances of failed transfer of technicians onto the
demands highly reliable wind turbines to curtail the turbine.
maintenance cost and to shorten the downtime. Power x Chances of technicians to be abandoned at the turbine
converter is one of the critical components that undergoes high location because of deterioration in weather conditions.
rate of medium and short term thermal cycles especially in
x Chances that the technicians will be unable to perform
Offshore floating wind turbines compared to fixed bottom
the maintenance intervention because of sea-sickness.
turbines. The current study proposes a novel methodology to
predict the remaining useful life of an offshore wind turbine Utility scale offshore turbines consists of large number of
power converter in digital twin frame work as a means of components that needs huge number of maintenance tasks. In
predictive maintenance strategy. The remaining useful life is order to minimize the O&M tasks and associated cost,
estimated for both diagnostic and prognostic health monitoring x The most urgent tasks must be prioritized
specific for offshore operating environment. x Vessels, crew and technicians needs to be effectively
utilized
Keywords-Remaining Useful Life (RUL); Physics based x Reduce the spare parts inventory
modelling; Power converter reliability; Wind turbine operation x Schedule preventative maintenance to minimize the
and maintenance (O&M) turbine downtime
I. INTRODUCTION An apt maintenance strategy is based on the nature of the
equipment and the conditions in which it operates. Lloyds
Offshore wind energy is gaining momentum for its Register is capable of modelling all the influencing variables
consistent and high wing energy potential compared to and provides the most advanced predictive maintenance
onshore wind. According to EWEA report, by 2016, more strategy.
than 12.631GW of offshore wind power is grid connected in
Europe [1]. About 25% of the Levelized Cost of Energy
(LCoE) is accounted for O&M [2] for the turbines that is
installed offshore or on remote locations, where the turbine
accessibility depends on the favorable weather conditions
(wind, waves, water levels, tidal currents, daylight, and
temperatures). Hence, reducing the O&M cost is of
paramount importance to lower the power generation cost
comparable to fossil fuel-based energy sources. Improving
the reliability of the critical components or predicting the
failures through improved maintenance strategy are the two
main approaches to reduce the cost of O&M and in turn
LCoE.
The dependency factors that increases the cost of
offshore power generation through increased O&M costs are Figure 1. Cost breakdown [3]

978-1-5386-5236-7/18/$31.00 ©2018 IEEE 197

 The objective of any maintenance strategy is to increase enrich wind farm operators and owners to decide on end-of-
the availability of the wind turbine for power generation. warranty review and its inspections, due diligence, life
Wind turbine manufacturers and operators prefer condition- extension, re-powering, retrofitting of major components and
based maintenance rather than primitive time based decommissioning. Physics based RUL prediction will greatly
preventive maintenance. The status and the condition of the influence those operations with accurate predictions.
components can be obtained through network integrated
Supervisory Control and Data Acquisition (SCADA) system. II. OFFSHORE WIND TURBINE POWER CONVERTER AND
The physics based remining useful life prediction together THERMAL FATIGUE LOADING CYCLES FOR RUL
with condition-based maintenance will be the best strategy ESTIMATION
for offshore wind turbines that is capable of lowering the As shown in Figure 3, power converter Insulated-Gate
labor cost as the labor cost alone contributes 35% of total Bipolar Transistor (IGBT) is constructed with multiple layers
repair cost as shown in Figure 1. of different thermal conductivity materials. In Direct Bonded
As shown in the Table 1, extracted from Energy research Copper (DBC) IGBT the die is attached to copper layer via
Centre of the Netherlands (ECN), an optimized O&M solder. For an electrical isolation, ceramic layer is placed in-
strategy will reduce repair costs by 10.43% and increases the between the copper layers. For high power applications such
profit by 4.21% with the net savings of EUR 1.06 million per MW scale wind turbines, base plate is typically attached to
year. air or water-cooled system. A few stacks of IGBT modules
will be used in generator/rotor side and grid side to form
TABLE I. O&M COST REDUCTION IN RESULTS OF O&M STRATEGY
OPTIMIZATION [3]-ECN- ECN-E--14-028 complete power converter system.

As risk assessment plays a vital role in O&M strategies

of offshore wind farms, a comprehensive failure mode and
effect analysis (FMEA) is carried out to determine critical
assemblies of a generic offshore wind turbine. As shown in
Figure 2, the most critical component is frequency or power
converter. A novel physics-based methodology has been
proposed to predict the Remaining Useful Life (RUL) of
electrical components especially the power converters of
variable speed wind turbines. The proposed methodology has Figure 3. Schematic diagram of full scale power converter and IGBT
been executed for both fixed and floating wind farm structural cross section view
applications by using 5MW NREL virtual wind turbine
along with digital twin technology In both onshore and offshore environment, wind speed
and temperature variation lead to intricate power loadings
(overload condition and system transients) on the power
converter and attributed for high failure rates. Literature on
power converter addressing the components and their
reliability are limited. The study [5] on power converter
identifies the components with high failure rate are IGBT
modules, converter control unit (CCU) and main fans.
Different CTE (coefficients of thermal expansion) of
adjacent materials layers in the IGBT are primary cause of
thermal driven failure during thermal cycles [6]. The
principal modes of failure are bond wire lift-off and solder
delamination or crack (at die to copper interface and copper
to baseplate interface).
A typical thermal cycle experienced by a wind turbine
power converter is shown in Figure 4. As most of the wind
turbines are not incorporated with condition monitoring for
power converter systems, a robust predictive method has to
be established to find the junction temperature, temperature
Figure 2. Critical assemblies ranked by risk priority number [4]
fluctuations and their influence on the RUL. With condition-
IEC 61400-28 - Through life management and life based monitoring system, it is much easier to predict the
extension has been proposed by the IEC wind committee to damage accumulation and RUL. Mission profile for each

198
category (long, medium and short term) thermal cycles is
different and the approach to predict the damage
accumulation is also varies accordingly.

Figure 6. Thermal network of power semiconductor devices for the long-

Figure 4. Wind power converter thermal cycles [7]. term thermal profile generation [7]

i) Thermal load due to environmental condition The existing wind farm data logging interval is 5 mins or
Thermal loads due to environmental conditions impose more and the total damage accumulation for this interval will
long term thermal cycles dictated by wind speed and ambient lead to substantial error if the long term thermal loading
temperature. By employing SCADA data, similar to 3 hour alone is examined. A predicted IGBT temperature is shown
interval period as shown in Figure 5, the damage in Figure 7 for long term thermal cycle of 5MW NREL wind
accumulation due to long term thermal cycles can be turbine. Since, medium and short term thermal loading
predicted. For floating wind turbines, relative wind speed is effects pose significant damage accumulation on the weak
used instead of freestream wind speed due to the platform points of (C Solder-Chip solder; B Solder-Base plate solder;
motion. bond wire) the IGBT structure.

Figure 7. Junction temperature and its fluctuation for wind turbine yearly
data

ii) Thermal load due to mechanical systems of wind turbine

Figure 5. One year mission profile of wind speed and ambient temperature Figure 8. Thermal network of power semiconductor devices for the
from wind farm/turbine (3-h averaged) [7] medium term thermal loading [7].

From the above wind and ambient profiles, loading In order to account the effect of thermal cycles less than
profile will be generated (in terms of torque and speed) from 3hour in the damage accumulation prediction, the loading
Aero-Elastic-Servo-Control model (for example FAST – profile with time constants of seconds to minutes is
NREL open source Aero-Elastic-Servo-Control code) for life established for each wind speed at 2m/s step [4] for the entire
time estimation. In a more convenient way the power loss wind spectrum of the whole year. Afterwards, the loading
will be estimated from the SCADA data power values to profile is generated using the Aero-Elastic –Servo code
predict the junction temperature and its fluctuation through a [FAST] and Miner rule is applied for the estimation of total
customized thermal model for damage accumulation and damage accumulation by medium thermal cyclic loading. By
RUL estimation. A simple thermal network model is used to this approach, the effect of pitch and speed control behaviors
compute the junction and case temperature for the IGBT can be recognized in damage accumulation prediction
module as shown in Figure 6. The values of the thermal process. Thermal capacitance is included in the thermal
network model parameters are provided by the power network model to capture the thermal transients as in the
converter manufacture. Figure 8.

199
iii) Thermal load due to electrical system FMEA (Failure Mode and Effects Analysis) to identify those
Short term thermal behavior is attributed to high failures that have significant impact on the wind turbine
frequency current oscillation in the electrical system leading operation and cost of maintainability. A generalized offshore
to junction temperature fluctuations. The oscillations are specific physics-based methodology is proposed by
smaller in amplitude and close to fundamental frequency of IDCORE-Lloyd’s Register research project to predict the
the power converter output. Cyclical amplitude is RUL of power converter of wind turbines as shown in Figure
analytically solved as shown in Equation [1]. 9.
Wind profile of the wind turbine is obtained from
͵ ͳ SCADA data. Then the profile is used in the integrated
οܶ௝ ൌ ܲ௟௢௦௦ ‫ܼ כ‬௧௛ ‫ כ‬൬ ൰ ‫ܲʹ כ‬௟௢௦௦ ‫ܼ כ‬௧௛ ‫ כ‬൬ ൰ሺͳሻ dynamics offshore wind turbine FAST model (Aero-elastic-
ͺ݂௢ Ͷ݂௢ servo-turbine model) which generates toque and speed for
the generator model input. A high level generator model
where ܲ௟௢௦௦ is the loss of power device; ܼ௧௛ is device thermal validation is possible by predicting power loss directly from
impedance; ݂௢ is the fundamental frequency of the converter SCADA data and uses it as an input to thermal model to
output. predict the junction temperature. Then, this can be validated
Previous studies indicate that, short term thermal loading with FAST and induction generator model generated power
significantly affects the total damage accumulation at rated loss. Once it is validated, analytical model of induction
wind speed conditions and hence it is compelling to include generator is considered to be validated. The flow diagram
for accurate prediction. All the three loading conditions are shown in Figure 9 is applicable for both diagnostic (to
used for the accurate prediction of damage accumulation and predict damage accumulation) and prognostic (to predict
RUL. Medium and short-term loading cycles are found to RUL) approach.
have greater influence on RUL prediction for offshore
floating wind turbines due to its control strategy. i) Integrated aero elastic servo control code
5MW NREL turbine FAST (aero-hydro-servo- elastic
III. PHYSICS BASED REMAINING USEFUL LIFE
simulator) model is used in this study. An accurate and
PREDICTION MEHODOLOGY
validated aerodynamic model (by Lloyds Register) for
As discussed in the previous sections it is found that the floating wind turbines is used in FAST. TurbSim code is
mean junction temperature and temperature cycling are the used to generate turbulent flow field for FAST. Various
main cause of temperature induced stress in the power FAST sub model files are generated by a Matlab based
converter assemblies that affect the reliability or life time automation. The details of the electrical drive are not
adversely. A few research studies resulted with power modeled in FAST; instead, the focus is on getting the torque-
converter models for damage accumulation prediction which speed curve as accurately as possible (which effects turbine
are still to be calibrated for its applicability for offshore wind loads). Though it is possible to extract the details of
turbine O&M applications i.e., for both fixed bottom and electrical drive from MATLAB Simulink generator models,
floating wind turbines. the most sophisticated built-in Thevenin-Equivalent Circuit
Until now statistical approaches are largely used for (TEC) model for induction machine available in FAST is
health diagnostics and prognostics of wind turbine employed.
components, as statistical approach does not completely
represent the real condition of the component. Due to the ii) Python based induction generator model
recent technology advancements in sensor and data-analytic A squirrel cage induction generator (SCIG) model is
approach use operational data such as CMS (Condition included to extract voltage and current variables to
Monitoring Systems) and SCADA to monitor the supplement FAST simulations through specific algorithm.
performance parameters. Physics-based modeling approach SCIG is a three-phase induction machine with three
accounts the failure mechanism and determines the system windings in the stator and three or more windings on the
condition using physical laws combined with operational rotor (assumed). Generators can be described with same set
data [8]. of equations similar to the motor, simplified to compute
voltage and current.
ii) Python Based Power Loss Prediction Model
There are three losses in IGBTs and Diodes in the power
converters. They are conduction losses (ܲ௖ ), switching losses
( ܲ௦௪ ) and blocking (or leakage) losses ( ܲ௕ ). Normally
blocking losses can be neglected. All the above losses can be
predicted accurately by developing a detailed mathematical
formulation from power converter data sheet or book. The
formulation can be coded in the Python.
Figure 9. RUL prediction flow diagram
ܲ௟௢௦௦ ൌ ܲ௖ ൅ ܲ௦௪ ൅ ܲ௕ ሺʹሻ
Prior to the physics based modeling, a detailed risk
assessment has to be performed by using comprehensive

200
iii) Python based thermal model for junction and case A framework is developed for physics based RUL
temperature prediction prediction and for a comparison between fixed and floating
The input for the thermal model is power loss. By which, offshore 5MW NREL wind turbines to quantify the effect of
the python code will predict the junction temperatures of changes in medium and short term thermal cyclic loads on
IGBTs and diodes together with case temperature also. the power converter. A schematic diagram of digital twin
Junction temperature Tjm and junction temperature technology platform is shown in Figure 11 for this
fluctuation dTj are given by the following equations 3 and 4: technology demonstration project.

ସ ସ

ܶ௝௠̴೅Ȁವ ൌ ܲ௟௢௦௦ ‫ כ‬෍ ܴ௧௛௝௖ି ் ൅ ሺܲ௟௢௦௦ ‫ כ‬෍ ܴ௧௛௖௔ିሺ௝ሻ

஽ሺ௜ሻ
௜ୀଵ ௜ୀଵ
൅ ܶ௔ ሺ͵ሻ
೟೚೙
೅
ഓ೟೓ೕ೎ష
ሺଵି௘ ವሺ೔ሻሻ̰ଶ
݀ܶ௝̴்Ȁ஽ ൌ ʹܲ௟௢௦௦ ‫  כ‬σସ௜ୀଵ ܴ௧௛௝௖ି்Ȁ஽ሺ௜ሻ ‫כ‬ ೟೛ (4)
೅
ഓ೟೓ೕ೎ష
ଵି௘ ವሺ೔ሻ

In (3), ܴ௧௛௝௖ is the thermal resistance from the junction to

case of the power module, ܴ௧௛௖௔ is the thermal resistance of
the air cooling, in which subscripts T and D denote IGBT
Figure 11. A simple schematic diagram of digital twin architecture for
and the freewheeling diode, whereas subscripts i and j denote 5MW NREL virtual fixed and floating wind farm O&M
four-layer and three-layer Foster structure for power module
and air cooling, respectively. ܲ௟௢௦௦ is the power loss of each For the first three major components as in Figure [4],
power semiconductor, and Ta is the ambient temperature. In virtual sensors are placed as shown in Figure 12 in the 5MW
equation (4), ‫ݐ‬௢௡ denotes the on-state time within each numerical turbine for predicting RUL for optimum O&M
fundamental period of the current at the steady-state strategy comparison, from which is derived the risk of failure
operation, ‫ݐ‬௣ denotes the fundamental period of the current, for each failure modes. Numerical models used during the
and τ denotes the thermal time constant of each Foster layer. wind turbine design phase (5MW NREL FAST model)
enable estimates to be made of degradation and remaining
iv) Python based rain flow counting method
life. Examples include replacing statistics of mean time
Rainflow counting method is employed to estimate the between failures and substituting with mean usage to failure,
thermal cycle frequency and amplitude [4]. The method was which may imply counting the number of use cycles for a
initially developed for material science and later applied to component or the fatigue damage due to thermal cycles.
power electronics. The local high and low in the data is
considered as peaks and valleys and the intermediate range is
assumed to be half cycles. The algorithm pairs the half
cycles to generate full cycles with respect to mean. The same
algorithm is developed in Python.
Figure 12. Virual sensor details in digital twin technology platform for
IV. PROPOSED APPROACH FOR DIGITAL TWIN PLATFORM optimal decision making of maintenance strategy

With the profound knowledge of Lloyds register’s Under the digital twin platform, virtual sensors have been
FMEA and RUL prediction experience with fixed bottom derived for locations in the structure without strain gauges
wind turbines, a detailed methodology is developed for using a combination of aero elastic models and finite element
offshore fixed and floating platforms. The Figure 10 shows a methods. The minimum requirement is to use wind speed,
brief detail about 5MW NREL numerical turbine is used in ambient temperature, yaw angle and electrical power
virtual space for digital twin technology framework. generated from SCADA records as shown in Figure 12. To
support smart maintenance planning and predictive
maintenance for wind farm operations, the following are the
areas under focus for fundamental research and the
development of new technology are applied within digital
twin framework:
1. Short-term forecasting at turbine locations – waves,
wind, currents, water depth
2. Estimating safe transfer windows – incorporating
simplified vessel response models
3. Statistical reliability of offshore fixed structures –
Figure 10. 5MW Wind turbine configuration for fixed and floating
application: Schematics of information mirroring model incorporating models of models

201
 4. Turbine in-service loading and accumulation of and wave together with mooring line failures. Well
damage – electronics, structures, mechanical systems, linked established Monte Carlo Markov Chain (MCMC) technique
with statistical reliability models and environmental forecasts is employed to generate futuristic event load profile
5. Technology for remote presence at difficult to- access sequences with higher confident level. Probability of
sites, wearable devices, streamlined exchange of data. occurrence of each load cases can be mapped from the
6. Synergies between workboat routes and bathymetric SCADA data similar to the one shown in Figure 14. The
surveys, local scour future sequence of events for DLCs computed from SCADA
7. Formal methods for quantitative risk assessment. data for next 5 hours is shown in Figure 15.
8. Safety culture, best practice, emergency preparedness,
design of systems and testing of procedures
The scope of the current study is limited to develop
physics based methodology for RUL prediction of power
converter (the type of converter is chosen from literature) by
considering all three types of thermal loads. A simple
process flow of the approach is given in the Figure 13 for
digital twin platform as a module to it.

Figure 14. Probabilistic occurrence of DLC load cases

Figure 13. Schematic diagram of RUL prediction process flow for power Figure 15. Example of future scenario generation by MCMC for 5 hours.
converter
IGBT junction temperature prediction for the next 5hrs is
For accurate power converter RUL prediction, it is
shown in Figure 16 by considering medium term thermal
necessary to create the similar environment experienced by
load in fixed bottom 5MW NREL turbine.
the turbine in terms of structural and electro- mechanical
behavior, and wind profile. For the current research, input
load cases for the FAST modelling are obtained through IEC
61400-3 [9]. As the power converter is subjected to
temperature cyclic load, fatigue-based design load cases
(DLCs) are chosen. The following design scenario is
considered for fixed bottom offshore wind turbines as shown
in Table 2.
TABLE II. FATIGUE LOAD CASE SCENARIOS AS PER IEC64100-3 FOR FIXED
BOTTOM OFFSHORE WIND TURBINES

Figure 16. IGBT temperature prediction for 5hrs

In the case of floating wind turbines, additional load Figure 17. Damage accumulation prediction for 5hrs based on IGBT
cases are to be included, accounting the interaction of wind temperature

202
 After IGBT and diode junction temperature prediction for generate the cycles counts and RUL estimation algorithm
the next 5hrs, rain flow counting algorithm is executed to will generate damage accumulation as shown in Figure 17.

Figure 18. Example of physics based RUL of power converter module output in digital twin platform

The physics based RUL prediction of power converter wind turbine/farm influences the power converter reliability
module outcome is shown in Figure 18 for fixed bottom models. Above prediction process can be expanded to wind
5MW NREL turbine. Digital twin framework includes the farm level as shown in Figure 19.
verification and validation process as and dynamic data of

Figure 19. Output for wind farm level case of power converter RUL module in digital twin platform for O&M strategy planning

During real time planning, the relevant information’s used presents as high-level performance indicators that influences
in algorithms should be presented in simple and clear form the decisions of O&M planning and execution. Figure 20
for making quick and optimal decisions. Digital twin shows the optimum decision-making process in wind farm
platform converts big data into manageable small data and project life cycle for O&M at wind farm level.

Figure 20. Optimal decision making by digital twin for O&M

V. SUMMARY AND CONCLUSION ACKNOWLEDGMENT

A physics based methodology is successfully Authors acknowledge the great support of Lloyds
implemented to predict damage accumulation and RUL Register to carry out the simulation work for this project.
accurately for fixed and floating offshore wind turbine power
converter by accounting medium and short term thermal REFERENCES
transient loadings together with long term thermal loading. [1] EWEA 2016 wind report , European Wind Energy Association,
Moreover, a complete framework is established for a https://www.ewea.org/
complete digital twin platform for optimum predictive [2] G. J. W. van Bussel, A. R. Henderson, C. A. Morgan, S. Smith, R.
maintenance strategy. Barthelmie, K. Argyriadis, A. Arena, G. Niklasson, and E. Peltola,
“State of the Art and Technology Trends for Offshore Wind Energy :
VI. FUTURE WORK Operation and Maintenance Issues,” in Offshore Wind Energy
Special Topic, 2001..
The future work includes the simulation of floating wind
[3] O&M Cost Reduction of Offshore Wind Farms - A Novel Case Study
turbine by imposing the boundary conditions of floating (https://www.ecn.nl/publications/PdfFetch.aspx?nr=ECN-E--14-028).
platform in order to quantify the difference in RUL [4] Marco Sepulveda, Peter Davies, Mark Spring, Dr Jonathan Shek, Dr
prediction between fixed and floating platform scenarios. Philipp Thies, Dr Erkan Oterkus, “Risk assessment of an offshore
wind turbine and remaining useful life (RUL) estimation of the power

203
 converter. Improving availability by prioritising failures with higher [7] Ke Ma, Marco Liserre, Frede Blaabjerg, and Tamas Kerekes,
risk to operation ( https://core.ac.uk/download/pdf/83924632.pdf). “Thermal Loading and Lifetime Estimation for Power Device
[5] Birk, J.; Andresen, B., "Parallel-connected converters for optimizing Considering Mission Profiles in Wind Power Converter”, IEEE
efficiency,reliability and grid harmonics in a wind turbine," European Transaction on Power Electronics, vol 30, no. 2, Feb 2015
Conference on Power Electronics and Applications, 2007, pp.1-7, 2-5 [8] Breteler D, Kaidis C, Tinga T, Loendersloot R. Physics based
Sept. 2007, doi:10.1109/EPE.2007.4417318 methodology for wind turbine failure detection, diagnostics \&
[6] K. Fischer, T. Stalin, H. Ramberg, T. Thiringer, J. Wenske, and R. prognostics. European Wind Energy Association; 2015;
Karlsson, Investigation of converter failure in wind turbines," in [9] IEC 61400-3:2009. Wind turbines – Part 3: Design requirements for
Elforsk report, Nov. 2012. offshore wind turbines.

204