Analytical Techniques for Performance Monitoring of Modern Wind Turbines

Presented at EWEA 2012 Copenhagen 16-19 April 2012
Analytical techniques for performance monitoring of
modern wind turbines
Staffan Lindahl
GL Garrad Hassan
[email protected]
Keir Harman
GL Garrad Hassan
[email protected]
Keywords: Wind Turbine Performance, Optimisation, Diagnostics, Changepoint
Abstract
There is strong demand in the wind industry
for understanding, in great detail, the
performance of operational wind farms. This
can be achieved by collecting sufficient data
from the standard SCADA system, the careful
management of those data and the use of
intelligent interrogation techniques.
Recent years have seen rapid development in
wind turbine technology, with modern turbines
adopting variable speed and pitch concepts
and complex solid state power conversion
technology. Also, increasingly intelligent
controller software programs now often include
self protection algorithms, noise reduction and
dynamic power curtailment as standard.
In line with these developments, the analysis
techniques commonly used in the industry,
which were primarily developed for more basic
technology in relatively small wind farms, have
become outdated. Such techniques no longer
reveal all of the attributes of modern wind
turbine performance and are also typically
difficult to implement on large wind farms in a
cost effective way.
This paper describes how SCADA data from
modern wind turbines can be analysed to
mitigate sub-optimal performance. The paper
describes which data signals tend to be the
most useful and what to look for. Moreover, it
describes how these signals can be monitored
by an automated system to ensure that
deviations from optimal performance are
detected rapidly, facilitating prompt corrective
action.
1 Introduction
The importance of understanding the
performance of an operating power plant is
clear. This applies not only to conventional
power generation but also to wind turbine
plants. Most modern wind farm SCADA
systems record the data required to meet this
need, but very few integrate the diagnostic
tools required to make full use of them. A
major challenge lies in analysing the
overwhelming volume of data gathered in a
timely and efficiency manner.
This paper presents a summary of typical
performance issues that affect modern wind
turbines and how these can be identified
through the analysis of SCADA data. It then
describes how automated algorithms can be
employed to detect these issues.
The techniques discussed are based solely on
the most generic data parameters recorded by
modern wind turbine SCADA systems, and
focus on:
• Assessing wind turbine availability;
• Identifying and diagnosing poor power
curve efficiency;
• Automation of wind turbine performance
analysis.
The techniques presented in this paper have
been developed by GL Garrad Hassan over
the past 20 years. They are based on a
detailed understanding of the engineering of
wind turbines and experience gained from
analysing SCADA data from over 30 GW of
operating wind turbine power plant from
around the world.
Presented at EWEA 2012 Copenhagen 16-19 April 2012
2 Assessing wind turbine
availability
There are several definitions of availability of
wind turbine plant and the meaning of the term
to a particular individual is often subjective.
For example, an electrical grid outage may be
considered available time by a wind turbine
manufacturer, whereas a wind turbine owner
may consider it unavailable time. When
looking at availability of operating wind turbine
plant from a diagnostics and optimisation
perspective, it is typically appropriate to
consider all causes of downtime, regardless of
the allocation of contractual responsibilities.
There are several different ways to use
10-minute SCADA data to define such a
measure of availability, here referred to as
‘Run-Time Availability’ (or ‘RTA’). While the
best options are often derived from system
specific counters, there are a number of
methods which can commonly be applied
generically to modern wind turbines.
The key considerations when devising the
RTA logic are to ensure that it distinguishes
between failure to operate due to low wind
speeds from failure to operate due to
malfunctions of the generating system and that
it captures 10-minute records when the turbine
was only partially available. Depending on the
application, it may also be necessary to
identify data affected by high wind speed shut-
down.
In the example shown in Figure 1, the
10-minute records affected by unavailability
have been identified by considering a
combination of power, wind speed, rotational
speed and pitch data. The key observations in
this example are that:
• Partial availability records have been
identified, appearing as black dots at
power levels above 0 kW;
• At wind speeds below the cut-in wind
speed, records where the turbine was
idling awaiting wind have been separated
from those where the turbine was faulted.
The percentage of down-time (i.e. 1-RTA) can
be evaluated from this as the count of
‘unavailable records’, as a fraction of the total
Figure 1: Identification of unavailability in 10-
minute SCADA data
number of data points, although care must be
taken to ensure fair weighting of the partially
available data points. This measure of RTA is
an estimate of the time-based availability of
the system during the periods of full SCADA
data coverage.
An important aspect to consider at this stage is
the distribution of the periods of downtime in
relation to wind speeds. In general, persistent
strong relationships between wind speed and
downtime are unusual [1, 2]. However when
external factors such as utility enforced
curtailment or structural load management
occur, or the availability is considered over a
relatively short operating period, say a month,
the effects can be significant. The
consequences of a positive correlation
between wind speed and downtime would be
higher energy losses than indicated by the
time-based availability measure.
One further aspect to evaluate is the
correlation of downtime and data loss. In the
authors’ experience, it is rare for SCADA
systems to have 100 % coverage of (valid)
data and the periods of lost data are often
related to downtime. An estimate of the
availability during periods of data loss can be
derived by inferring the amount of production
unaccounted for in the incomplete SCADA
data set, by comparing it to a dataset with full
coverage. Typically, the production measured
by the utility at the point of interconnection with
Presented at EWEA 2012 Copenhagen 16-19 April 2012
the electrical grid can be used. Additional
prerequisites for this analysis are:
• An appropriate estimate of the wind farm
internal electrical efficiency;
• It can reasonably be assumed that the
wind speeds over the period of full data
coverage are representative of the periods
of data loss; and
• An appropriate estimate of the energy to
time weighting of the RTA for the period of
data loss.
Applying these assumptions yields an estimate
of the availability during the period of SCADA
data loss. Availability measures representing
100 % SCADA data coverage can then be
derived, either on the basis of time or energy.
It should be noted that, when considering
other aspects of availability, where perhaps
accountability and contractual obligations are
of interest, or reliability is the main focus, a
different approach to classification of
downtime events is required.
The International Electrotechnical Commission
(IEC) provides recommendations for the
classification of downtime events in the
technical specification IEC 61400-26 part 1. In
order to classify turbine downtime events
retrospectively using this model, information
beyond that held in most SCADA systems
(such as Operation and Maintenance (O&M)
records and service reports) will typically be
required. Through this method, downtime
events can be classified by accountability.
SCADA fault data can be used in conjunction
with turbine taxonomy-based mapping of
downtime events, as per the ‘Reliawind
Project’ [3], to derive the reliability profiles of a
set of wind turbines, as shown in Figure 2,
reproduced from [3]. Through this process,
energy losses incurred can be attributed to the
failure of individual turbine sub-systems. This
approach can be extremely valuable from an
optimisation perspective.
3 Diagnostics of power cure
performance
The identification and diagnosis of the causes
of performance issues requires careful
interrogation of SCADA databases, often in
conjunction with reviews of service records
and O&M reports. For certain issues, it may
even be necessary to inspect components of
the turbine physically, in order to successfully
diagnose the root cause of failures.
Figure 2: Wind turbine reliability profiling [3]
Presented at EWEA 2012 Copenhagen 16-19 April 2012
Figure 3: Impact of wind vane misalignment on wind turbine performance
There are many causes of changes in the
power curve performance of wind turbines:
pitch control malfunction, blade damage and
fouling, software updates, aerodynamic
enhancements, controller malfunction, blade
angle calibration, sensor misalignment, self-
protection and noise reduction modes - to
name just a few. Each of these issues can be
characterised as being either known or
unknown to the turbine software controller.
This is a key characterisation, as it affects the
diagnostics approach. Typically, issues known
to the controller will be apparent in the SCADA
data as inconsistencies in the power vs. torque
or the power vs. pitch relationship. This is
typically not the case for issues unknown to
the controller.
The example shown in Figure 3 illustrates the
effect of wind vane misalignment on the power
curve performance of a wind turbine. No clear
associated changes in torque and pitch
characteristics are apparent. In order to
confirm that such performance issues are
indeed real and not the result of wind speed
measurement inconsistencies, it is often
necessary to compare the wind speed and
power data of the subject turbine to a nearby
reference point (e.g. cross-correlation to a
meteorological mast or another wind turbine).
Care must be taken to ensure that the
reference data are consistent over the entire
test period and careful consideration of the
uncertainties in the correlations used is
required.
Further cross-correlation and time-series
analysis of the wind or nacelle direction data
will, in general, reveal if the cause of the
reduction in performance can reasonably be
attributed to wind vane misalignment, although
it is often necessary to inspect the
instrumentation in order to confirm the cause
conclusively. Performance will typically return
to normal levels once the instrumentation has
been realigned.
While the main focus of the diagnostics
assessment may be to understand the power
curve performance of the turbine, examination
of wind speed and power data is often not
sufficient. Many issues cause only subtle
changes in power curve performance, but will
over long periods of time still cause significant
energy losses. Typically, such issues may
arise from software updates or changes to
parameters in the turbine software controller
aiming to, for example, control noise levels or
reduce loading on the wind turbine structure.
The example in Figure 4 shows significant
changes in the power vs. torque
characteristics of a turbine for an operating
period of five years. A review of power curve
data alone is unlikely to have revealed this
issue, as is demonstrated in Figure 5.
Even though the change in power curve
performance in this example was very small,
issues such as this one may go undetected for
long periods of time, incurring a significant
energy loss. In this particular example, it
represented an energy loss of approximately
Presented at EWEA 2012 Copenhagen 16-19 April 2012
Figure 4: Power vs. torque over a 5-year
operating period
Figure 5: Power curve data for the 5-year
operating period shown in Figure 4
1 % on an annual basis and affected the entire
wind farm for over a full year of operation.
Such a change in energy yield can have a
substantial impact on the financial
performance of the project [4, 5].
4 Automation of wind turbine
performance analysis
As seen in the previous sections, the
challenge of a diagnostics assessment of wind
turbine performance is largely one of
identifying changes in the relationship between
SCADA data signals. The analysis does
however require commitment, time and the
use of sophisticated software tools combined
with the knowledge of an experienced
engineer [4]. This can make it an unfeasible
undertaking for many stakeholders in the wind
energy industry. A technique for automating
this process, thereby removing those barriers,
has therefore been developed.
4.1 Multi-channel Changepoint
Analysis
A method based in Changepoint Analysis has
been developed by the authors for the purpose
of identifying any changes that lie within a
time-series of wind turbine operational data. In
order to ensure compatibility with a wide range
of data sources, a non-parametric cumulative
sum of error approach has been used [6]. To
balance processing overheads against
accuracy, a Binary Segmentation search
algorithm has been used [6, 7].
The Changepoint search algorithm is applied
to the residuals of the correlation of any two
concurrent time-series available in the SCADA
data of a modern wind turbine. While the
analysis has no prerequisites for which data
signals are used, intelligent decisions will aid
the analysis. Power vs. wind speed, power vs.
torque and blade pitch angle vs. wind speed
are all considered to be suitable combinations
and the technique can be extended to the
monitoring of temperature and vibration data.
The changepoint analysis can be broken down
into the following step-by-step procedure:
1. Establish a relationship of interest (e.g.
power vs. torque, pitch vs. power, etc.).
2. Establish the average relationship
between the two channels over the
analysis period. This can either be a
function of a suitable order or a simple
‘look-up table’.
3. Calculate the residuals (“∆”) of the actual
values, relative to the expected values,
given the relationship established in
Step 2.
4. Calculate the cumulative sum of ∆ (“ Σ(∆)”)
and for several (“i”) random permutations
(bootstraps) of ∆ (“ Σ(∆
r
)
i
”)
Presented at EWEA 2012 Copenhagen 16-19 April 2012
5. Determine the extreme range of Σ(∆) and
Σ(∆
r
)
i
, i.e:
Eq. 1
Eq. 2
With a null hypothesis (“h
0
”) that there was no
change in the time-series, the probability (“P”)
that a change occurred is defined by:
Eq. 3
The timing of the main change event is defined
by the temporal location of the Dominant
Extrema (“DE.”) of Σ(∆).
Figure 6 shows the results of applying this
search algorithm to the power and torque data
shown in Figure 4.
Figure 6: Changepoint analysis of 5-year time-
series of power vs. torque data
The blue data series shows the cumulative
sum of the residuals of the chronological time
series data (i.e. Σ(∆)). Positive gradients on
this line indicate a period where the data tend
to lie above the mean of the entire period.
Conversely, negative gradients indicate a
period where the data lie below the mean of
the period. Consequently, inflection points
indicate that a change in the trend has
occured. The statistical significance of the
change is determined, based on the
bootstraps.
The probability exceedance levels that h
0
should be rejected have here been calculated
for the 50 %, 75 %, 90 % and 99 % cofidence
levels. In this example, there is a probability
exceeding 99 % that a change occured
approximately one quarter into the dataset,
marked by the vertical dashed line in Figure 6.
It is important to note that this probability
applies only to the single data point where the
change was detected. A visual inspection of
Σ(∆) (i.e. the blue data series in Figure 6),
indicates that there may be other significant
changes within the data set.
Multiple changes within the data set are
handled by the Binary Segmentation method
[6, 7]; Essentially, when a change has been
detected, the data set is bisected at the
temporal location of the Dominant Extrema.
The analysis is then re-iterated for the two
resulting subsets of data and, if further
changes are detected, additional bisects are
applied until no further changes are detected.
Consequently, this method does not explore
every solution to the changepoint problem.
Alternative exact algortihms exist [7], but come
at a significant computational overhead when
handling large data sets.
Figure 7: Example of power curve affected by
de-rating
In performance monitoring applications,
detecting changes soon after first occurrence
is of paramount importance. The method
described here is suitable for this application
also, but exact methods for detecting multiple
changes, such as those proposed in [7], may
be more attractive alternatives due the
reduced concern with computational
overheads when dealing with small data sets.
Presented at EWEA 2012 Copenhagen 16-19 April 2012
Figure 8: Application of Changepoint Analysis to the de-rating example in Figure 7
Figure 9: Results of Changepoint Analysis for the last three days of data from Figure 8
Figure 7 shows a plot of power vs. nacelle
anemometer wind speed for an operational
period of approximately 31 days. A short
period of de-rating that affected the last
10 hours (approximately) of operation has
been highlighted. The results of applying the
Changepoint Analysis to this relatively short
time series of data is shown in Figure 8.
The analysis has detected a change very
close to the end of the data set, at a
confidence level somewhere between 90 %
and 99 %. A closer inspection of the results, in
combination with the raw power data is shown
in Figure 9. The method’s ability to pinpoint the
timing of the change and promptly after the
first occurrence is demonstrated.
4.2 The automated monitoring
system
In order to make use of Changepoint Analysis
in an automatic performance monitoring
system, a live or semi-live feed of SCADA data
is required. The system periodically analyses
pre-set data signal combinations for changes.
On detection of a change, warnings are issued
to the operator and/or owner of the plant. In
addition to information on the turbine affected,
the warnings can include information such as
confidence level, severity, how the change
was detected and suspected cause. Live
provision of this information would facilitate
prompt corrective action and minimisation of
energy losses.
Further benefits can be gained through
coupling the notifications with short-term
meteorological forecasting. Such a system
would allow an estimate of the energy loss that
will be incurred, if the issue remains
unresolved, to be made. Provision of
information at this level of detail to the wind
farm operator facilitates service scheduling
and prioritisation with a focus on energy loss
minimisation.
It is estimated that prompt resolution of all
power curve performance issues, such as
those discussed in this paper, could increase
the annual energy yield of operating wind
turbines by, in the order of, 1 % or 2 % [9].
Typically, in terms of Return on Investment
(ROI) for the project, a 1 % increase in energy
yield, and thus revenue, can result in an
increase in project profits by 10 % [1, 5].
Presented at EWEA 2012 Copenhagen 16-19 April 2012
4.3 Limitations and further work
Performing non-parametric Changepoint
Analysis on large datasets is computationally
intensive. Computing power, however, is
relatively cheap and this is therefore not seen
as a major limitation. Nevertheless, parametric
options should be explored to determine the
scope for using more computationally efficient
methods.
The use of exact search algorithms should be
explored further to improve the accuracy of the
search. This is particularly important for
accurately detecting small changes and
changes with a short duration.
The method used assumes an independent
error structure of the input data under the null
hypothesis. This is known to be a poor
assumption for some of the input data signals
due to the autoregressive effects of many
natural phenomena, such as wind direction
and air density. Consequently, the current
approach yields an elevated false positive
detection rate. There are several well
established methods for dealing with this, such
as those proposed in [10]. Research is
currently being conducted to include
corrections for autoregressive effects.
5 Conclusions
There are indisputably good reasons for
understanding and optimising the performance
of operating wind turbine plants. Most SCADA
systems for modern wind turbines record the
data required, but few provide the tools
required to make full use of them.
It has been demonstrated how the SCADA
data recorded at largely all modern operating
wind turbines can be used to identify and
diagnose the poor performance of wind turbine
plant and how even small changes in
performance can have significant impact on
the project finance. Performing the diagnostics
assessment regularly on large fleets of wind
turbines is however often unfeasible due to
time constraints and the lack of suitable
software tools, expertise and experience.
A performance monitoring algorithm has been
developed in order to remove these barriers It
has been shown that the method can detect
changes in the performance characteristics of
wind turbines promptly and reliably. It has
been described how this method can be
incorporated into an automatic monitoring
system, facilitating prompt resolution of issues
and minimisation of energy losses.
Based on the authors’ extensive experience of
analysing the performance of operating wind
turbine plants on a commercial basis, it is
estimated that the implementation of automatic
performance monitoring systems could
increase annual energy yields of typical wind
farms by 1% to 2 %, and ensure that the
projects consistently operate as designed.
There is, however, scope for improvement to
the method developed, particularly with
respect to the use of exact search algorithms
and normalisation of autoregressive effects in
the input data.
References
[1] Harman, K., Walker, R. And Wilkinson, M.,
“Availability trends observed at operational
wind farms”, Garrad Hassan and Partners Ltd,
European Wind Energy Conference, Brussels,
2008.
[2] Wilkinson, M., van Delft, T., and Harman,
K., ”The effect of environmental parameters
on wind turbine reliability”, European Wind
Energy Association, Copenhagen, 2012.
[3] Wilkinson, M., et al, ”Measuring wind
turbine reliability – Results of the Reliawind
project”, European Wind Energy Association,
Brussels, 2011.
[4] Harman, K. “Wind plant optimisation: Peak
performance”, Energy Engineering magazine,
Issue 20, 2008.
[5] Lindahl, S., Harman, K., Graves, A.M.,
“Optimising performance and asset return
through forensic analysis of SCADA data”,
American Wind Energy Association, 2009.
Presented at EWEA 2012 Copenhagen 16-19 April 2012
[6] Taylor, Wayne A. (2000), "Change-Point
Analysis: A Powerful new tool for detecting
changes," WEB:
http://www.variation.com/cpa/tech/changep
oint.html.
[7] Killick, R., Fearnhead, P. and Eckley, I.A.
“Optimal detection of changepoints with a
linear computational cost”, Jul 12 2011.
[8] Killick, R., Jonathan, P. and Eckley, I.A.
“Efficient detection of multiple changepoints
within an oceanographic time series”, 2011.
[9] Harman. K., “Availability and operating
efficiency – How well do wind farms actually
perform?”, Renewable UK 2010, Glasgow,
2010.
[10] R I Harris, “The macrometeorological
spectrum – a preliminary study”, Journal of
Wind Engineering and Industrial
Aerodynamics 96 (2008) 2294-2307.
[11] K D Harman & P G Raftery, “Analytical
techniques for understanding the
performance of operational wind farms”,
EWEC 2003.