Professional Documents
Culture Documents
Applied Energy
journal homepage: www.elsevier.com/locate/apenergy
h i g h l i g h t s
AFTCC schemes using an integrated FDD and FTC approach are developed.
A model-free FDD system is designed using a rule-based threshold test technique.
A model-based FDD system is designed using a fuzzy modelling technique.
Different simulations are performed on a high-fidelity offshore wind farm model.
Effective fault detection, diagnosis and accommodation results are achieved.
a r t i c l e i n f o a b s t r a c t
Article history: Given the importance of reliability and availability in wind farms, this paper focuses on the development
Received 18 August 2016 of fault diagnosis and fault-tolerant control schemes in a cooperative framework (referred to as active
Received in revised form 20 December 2016 fault-tolerant cooperative control) at the wind farm level against the decreasing power generation caused
Accepted 22 December 2016
by turbine blade erosion and debris build-up on the blades over time. In more details, the paper presents
Available online 6 January 2017
a novel integrated fault detection and diagnosis and fault-tolerant control approach oriented to the
design and development of two active fault-tolerant cooperative control schemes for an offshore wind
Keywords:
farm. Each of the schemes employs a fault detection and diagnosis system to provide accurate and timely
Fault detection and diagnosis (FDD)
Fault-tolerant control (FTC)
diagnosis information to be used in an appropriate automatic signal correction algorithm for accommo-
Fault-tolerant cooperative control (FTCC) dation of faults in the wind farm. The effectiveness and performance of the proposed schemes are eval-
Wind turbines uated and compared using different simulations on a high-fidelity offshore wind farm benchmark model
Wind farm in the presence of wind turbulences, measurement noises and realistic fault scenarios.
Turbine blade erosion Ó 2016 Elsevier Ltd. All rights reserved.
Debris build-up faults
1. Introduction limited accessibility and harsh climate conditions come into play,
resulting in higher failure rates and maintenance challenges. This
As one of the most cost-competitive forms of renewable energy, motivates the design and development of advanced fault detection
wind has remarkable potential for fulfilling the increasing demand and diagnosis (FDD) and fault-tolerant control (FTC) schemes in
for global energy in an environmentally responsible way. To fur- wind farms to improve their reliability and availability.
ther reduce the average cost of wind energy, large wind turbines The FTC schemes can be designed in either passive or active
are often installed in clusters called wind farms, particularly at off- ways. A passive FTC (PFTC) scheme employs the robustness of
shore locations. As more and more offshore wind farms are devel- the closed-loop control system to accommodate faults, while an
oped further from shores, both the factors of complexity and active FTC (AFTC) scheme reconfigures the closed-loop control sys-
tem after fault occurrence. AFTC schemes usually require FDD
information in the process of control reconfiguration.
q
This work was supported in part by the Natural Sciences and Engineering In general, the FDD and FTC schemes can be applied at both the
Research Council of Canada (NSERC) and Concordia University. individual wind turbine and entire wind farm levels. Recently,
⇑ Corresponding author at: Department of Mechanical and Industrial Engineering, research has been more focused on the application of such meth-
Concordia University, Montreal, Quebec H3G 1M8, Canada.
ods at wind turbine level (for example, see [1–5]). Most of these
E-mail addresses: h_bad@encs.concordia.ca, h.badihi@gmail.com (H. Badihi),
Youmin.Zhang@concordia.ca (Y. Zhang), Henry.Hong@concordia.ca (H. Hong).
works try to address the FDD and FTC problems in two standard
1
During his Sabbatical Leave from Concordia University. wind turbine benchmark models presented in [6,7]. A recent
http://dx.doi.org/10.1016/j.apenergy.2016.12.096
0306-2619/Ó 2016 Elsevier Ltd. All rights reserved.
H. Badihi et al. / Applied Energy 201 (2017) 284–307 285
review of the literature in [8] provides more references on FDD and to the other relevant works in the existing literature, it is worth
FTC for wind turbines. In reality, some faults are more easily mentioning that the proposed schemes in this paper are designed
detected, diagnosed, and accommodated at the wind farm level. and developed in a way to be valid for any layout of a wind farm
This can be performed by comparing the performance of turbines with any direction of the wind, while the considered fault may
operating under similar wind conditions. The FDD and FTC at the occur simultaneously in more than one turbine in the farm. This
wind farm level is a very recent field of research with only a few is actually a more general case that may happen in a real wind farm
research works reported in the literature. Most of these works in operation, and has never been studied in the available literature
are focused only on condition monitoring and fault detection in so far.
wind farms. In [9,10], various data-mining algorithms are applied The effectiveness and performance of the proposed AFTCC
to develop models for predicting possible faults in wind farms. In schemes are evaluated and compared using different simulations
[11], the relationship between the wind speed and the generated on a high-fidelity offshore wind farm benchmark model in the
power in a wind farm is estimated using three different machine presence of wind turbulences, measurement noises and realistic
learning models. The models can detect anomalous functioning fault scenarios. Moreover, extensive Monte Carlo simulations are
conditions of the wind farm, but they are unable to isolate and performed to evaluate the robustness of the proposed schemes
identify faults. More recently, researchers have studied the FDD with respect to modelling errors, disturbances and measurement
and FTC problems in a standard benchmark model presented in uncertainties.
[12] that represents a wind farm with nine turbines that are sub- The remainder of the paper is organized as follows: In Section 2,
ject to different fault scenarios. For example, the authors in [13] the wind farm benchmark model used in this paper is briefly
present a fault detection and isolation approach based on a set of described. The considered fault is described and analysed in
piecewise affine Takagi-Sugeno models that are identified from Section 3. Section 4 presents the AFTCC based on integrated FDD
noise-corrupted measurements. Borcehrsen et al. [14] present a and FTC approach against the fault discussed in Section 3. The
fault detection system relying on dynamical cumulative sum for details of the FDD at wind farm level are presented in Section 5.
residual evaluation, and a load distributing controller for accom- Section 6 presents the simulation results with some comments
modating possible faults. A passive FTC scheme is presented in and discussions. Finally, conclusions are drawn in Section 7.
[15] that integrates a fault estimation scheme with the design of
a controller accommodation system. An evaluation study is also
presented in [16]. Duviella et al. [17] propose an evolving classifi- 2. Overview of the wind farm benchmark model
cation algorithm for detection and isolation of faults due to debris
build-up on the wind turbine blades. In [18], the fault diagnosis is This paper considers an advanced wind farm simulation toolbox
conducted using interval nonlinear parameter-varying parity equa- called SimWindFarm that is developed as a part of the EU-FP7
tions, assuming an unknown but bounded description of the noise project, AEOLUS [21]. The toolbox provides a realistic wind farm
and modelling errors. Another work reported in [19] presents an simulation benchmark model that allows control designers to
active fault-tolerant control scheme based on a model-based FDD
approach. In actuality, the above cited research works are similar
in two aspects: first, they have assumed that only one fault occurs
at a time in a farm, and second, they mostly rely on wind speed or T8 T9 T10
its estimation which normally depends on the layout of the wind
D3
D2
farm and direction of the wind as well. For example, the algorithms
Wind
D2
paper presents a novel integrated FDD and FTC approach in a coop-
erative framework referred as the active fault-tolerant cooperative
T1 T2 T3 T4
control (AFTCC) by recognizing the differences in controlling a
wind farm from a single wind turbine. Here, the term ‘‘active” is D1 D1 D1
due to the integrated design of both FDD and FTC [20], while the
term ‘‘cooperative” is due to the cooperative design for multiple Fig. 1. Wind farm layout (D1 = 600 m, D2 = 500 m, D3 = 300 m).
wind turbines (wind farm) which is beyond a design for a single
wind turbine and will be demonstrated in the later parts of this
paper for both FDD and FTC strategies. The proposed approach is
oriented to the design and development of two AFTCC schemes
for an offshore wind farm against decreased power generation
faults that is caused by turbine blade erosion and debris build-up Wind Turbines Wind Field
on the blades over time. The first scheme is based on a model-
free FDD system that incorporates a rule-based threshold testing
technique for residual evaluation. Conversely, the second scheme
is based on a model-based FDD system that incorporates data-
driven models developed using a fuzzy modelling and identifica-
tion (FMI) technique. Both schemes are relying on an appropriate
automatic signal correction (ASC) algorithm that employs the pro- Wind Farm Network
vided accurate and timely FDD information for accommodating the Controller Operator
possible faults in a wind farm.
The proposed AFTCC schemes not only provide necessary FDD
information for condition monitoring purposes, but also provide Fig. 2. Illustration of overall model structure for N turbines. Note that the bold
the effective possibility of the accommodation of faults in a wind letters stand for sets of variables, for example P d ¼ ½P d;q with q ¼ 1; 2; . . . ; N (This
farm. To further highlight the contribution of this paper compared figure is based on [21]).
286 H. Badihi et al. / Applied Energy 201 (2017) 284–307
and applied in the benchmark model. Fig. 1 shows the default lay-
Wind Turbine # q
out for the considered wind farm with ten turbines. The overall
Local Wind structure of the simulation benchmark model under consideration
Profile is illustrated in Fig. 2.
As it is shown in Fig. 2, the wind farm simulation benchmark
Baseline model is composed of four major components in the top level that
Control operate in a closed loop:
System Network Operator: The network operator is responsible for
determining the total active power demand P D (also called opera-
tor’s total demanded power) required for a reliable connection of
the wind farm to the electrical grid. This can be performed in dif-
Fig. 3. The qth wind turbine in the farm (q ¼ 1; 2; . . . ; N). Note that in addition to
ferent modes, such as absolute, delta, and frequency regulation
the generated power P g;q , the model provides many other measured variables such
as coefficient of thrust C T;q , and so on. modes. This paper employs the frequency regulation mode, in
which the measured grid frequency f m is used as a feedback signal
to set up active power control in real-time. The objective is to main-
tain the necessary balance between power generation and load,
develop, implement, and investigate farm level control and diagno- which in turn regulates the grid frequency to its reference value
sis algorithms under different operating conditions for an optional f r , despite having a changing grid load. The baseline model for
quantity and layout of turbines installed in a wind farm. In the the network operator in the frequency regulation mode includes
benchmark model, sensor models are updated to represent a dead-band proportional gain control which employs frequency
noise-contaminated, uncertain measurement systems. The recom- error f e ðkÞ in (1) to determine the total demanded power PD ðkÞ at
mended rate and magnitude limiters are also applied on any refer- the time-step k in (2). Such a simple control scheme regulates
ence signal to the actuator models. In addition, to facilitate the the grid frequency to any specified reference value, for example,
assessment of the robustness features of any control solution 50 Hz in large areas of the world.
under external disturbances, different wind fields with arbitrary
mean wind speeds and turbulence intensities can be generated f e ðkÞ ¼ f m ðkÞ f r ð1Þ
Time
50 150 250 350 450 550 650 750 850 950
[S] 0 100 200 300 400 500 600 700 800 900 1000
Fig. 4. Timeline for the occurrence of the considered fault in a designated number of wind turbines in the farm. Note that, the total simulation time is 1000 s, and T# stands for
wind turbine number # with respect to the wind farm layout shown in Fig. 1.
5 5
Fault-Free Operation Fault-Free Operation
Generator Power [MW]
4 4
3 4.2 3
4
2 2
3.8
4
1 3.6 1
250 300 350 400 500 600 700 800 900 1000
0 0
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(a) (b)
Fig. 5. Generator power responses during fault-free and faulty operations of wind turbines: (a) T1, and (b) T3 in the farm.
H. Badihi et al. / Applied Energy 201 (2017) 284–307 287
Pa;q ðkÞ
Pd;q ðkÞ ¼ PD ðkÞ ; q ¼ 1; 2; . . . ; N ð5Þ
PA ðkÞ
Wind Turbines: As shown in Fig. 2, this component is devoted to
the simulation of dynamics of N turbines installed in the wind farm
FDD based on their measured nacelle wind speeds V nac ¼ ½V nac;q , effec-
Supervisor tive wind speeds V rot ¼ ½V rot;q , and power demands P d ¼ ½P d;q with
System
q ¼ 1; 2; . . . ; N. With respect to the outputs, the component gener-
ates a set of outputs including a set of measurements Mes required
for use by the wind farm controller, as well as a set of coefficients
of thrust C T ¼ ½C T;q for turbines which are necessary to calculate
the wake effects (i.e., low speed turbulent air flows behind turbine)
Wind Farm Network by the wind field component. In the component of wind turbines,
Controller Operator each turbine is simulated using a simple model of an offshore
5 MW turbine that has already been proposed by the U.S. National
Renewable Energy Laboratory (NREL) (see [22]). The baseline con-
Fig. 6. Schematic of the proposed AFTCC based on integrated FDD and FTC trol system used in each individual wind turbine is shown in Fig. 3.
approach. Here, Mes is a vector of performance data including measured variables As shown in the figure, the control system acts upon the power
and control commands/references in wind turbines. Based on FDD information IðkÞ, demand Pd;q in (5). This control system is essentially composed of
the supervisor applies appropriate signal correction/modification using power loss
^ ¼ ½P
estimates P ^ q . Note that the farm includes N turbines (q ¼ 1; 2; . . . ; N). a blade-pitch controller and a torque controller to compute the
appropriate reference blade-pitch angle br;q and reference genera-
8 tor torque sr;q , respectively. The blade-pitch controller is funda-
>
> 0:5ðP1 Þ d 6 f e ðkÞ 6 d mentally a Proportional-Integral (PI) controller that tracks a
>
>
>
> 0:5ðP1 P 2 Þ f e ðkÞ P c constant generator speed (rated generator speed) so that the tur-
>
>
< 0:5ðP þ P Þ f e ðkÞ 6 c bine operates at its rated power in the full-load region. The torque
1 2
PD ðkÞ ¼ ð2Þ controller is designed to both optimize power capture in the
> 0:5ðP Þ 0:5ðP Þ f e ðkÞd
> d < f e ðkÞ < c
>
> 1 2 partial-load region, and to improve output power quality in the
>
>
cd
>
> full-load region by varying the generator torque. More precisely,
: 0:5ðP1 Þ 0:5ðP 2 Þ f e ðkÞþd c < f e ðkÞ < d
cd the torque controller is set to be active for varying the torque, dur-
In (2), the constants c and d are, respectively, the control band ing both the below and above rated wind speeds. A more complete
and the dead band defined by user (c > d). Moreover, the power description of the wind turbine benchmark model is available in
parameters P1 and P2 are defined in (3) and (4), respectively. Here, [22].
P min and Pmax denote the prescribed minimum and maximum limits Wind Field: The wind field model represents the interactions
for the total power generated by the wind farm, respectively. between the wind turbines installed in a wind farm. This model
includes models of ambient fields and wake phenomena for simu-
P1 ¼ Pmax þ P min ð3Þ lating wakes meandering behind turbines and their effects on the
ambient wind field. Thus, the wind speed throughout the farm is
P2 ¼ Pmax P min ð4Þ obtained.
Wind Farm Controller: As it is shown in Fig. 2, the wind farm
controller acts as an interface between the network operator and 3. Blade erosion/debris build-up fault
wind turbines. Its main functionality is to ensure the appropriate
distribution of the operator’s total demanded power PD among In essence, decreased power generation in a wind farm may be
wind turbines in the farm. It also provides an estimate of total due to several different malfunctions. However, blade erosion
available power PA in the wind farm to the operator (e.g., in the along with debris build-up on the blades due to dirt, ice, etc. con-
case of delta mode operator). The baseline wind farm controller stitutes the most probable fault which results in a lower power
Table 1
Formation of R modules for a wind farm with N turbines.
generation because of changes in the aerodynamics of the wind a fault in a timely and effective manner. However, it is difficult to
turbine, and thereby lowering the maximum limit of obtained handle this fault at a wind turbine control level, mainly due to the
power. In more detail, the aerodynamic torque saer applied to the fact that a lower generated power may be due to either debris
rotor by the wind is defined in (6), in which xrot is the rotor angu- build-up on the blades or simply that the true wind speed is lower
lar speed, q is the air density, A is the swept area of the turbine than the measured/estimated wind speed. Conversely, at a wind
rotor, and V w is the wind speed [23]. farm level, it is possible to compare the performance and operation
features of the different wind turbines in a given wind farm.
1
saer ðtÞ ¼ qAV 3w ðtÞC p ðbðtÞ; kðtÞÞ ð6Þ The described benchmark model in Section 2, in its original
2xrot ðtÞ form, does not include any fault. However, it is possible to fairly
model a realistic scenario for the blade erosion and debris build-
In Eq. (6), the power coefficient C p ðb; kÞ is a three dimensional
up fault and incorporate it into the benchmark model. In the fol-
surface as a function of the tip-speed ratio k and the blade pitch
lowing part of this section, the modelling of the considered fault
angle b in which the latter two terms determine the operating con-
in this paper is described.
dition of a variable speed wind turbine. For variable-speed wind
In general, the generator power P g can be expressed as:
turbines, the turbine is ideally operated at the peak of the C p sur-
face in order to capture as much power as possible. However, over Pg ðtÞ ¼ gg gm P aer ðtÞ ð7Þ
time, the blade erosion and debris build-up fault shifts the tur-
bine’s C p surface downward, resulting in a lower energy capture in which gg is the generator efficiency, gm is the efficiency of trans-
through not only a decrease in the peak value of the C p surface mission system, and P aer is the aerodynamic rotor power given by
but also a change in the location of the peak of the C p surface. [23]:
The sensitivity of energy loss to the changes at the peak of the C p
Paer ðtÞ ¼ saer ðtÞxrot ðtÞ ð8Þ
surface caused by the aforementioned fault is considered and
quantified in [24], which concludes that the fault can lead to a sub- Now, substituting (8) into (7) and using (6) yields:
stantial non-optimal operation and power loss. Given the signifi-
cance of the blade erosion and debris build-up fault in wind ðgg :gm Þ
Pg ðtÞ ¼ qAV 3w ðtÞC p ðbðtÞ; kðtÞÞ ð9Þ
turbines, it is necessary to detect, diagnose, and accommodate such 2
DM1
(Turbine T1)
M1,2 DM1
(Turbine T1)
M1,3
M1,N-1
M1,N
DMZ
M2,3 DMZ (Turbine TZ)
(Turbine TZ)
M2,N
DMN
(Turbine TN) DMN
(Turbine TN)
MN-1,N
Decision Making
(DM)
FDD System
Fig. 7. FDD system including R modules each for conducting the monitoring of the consistency of the powers generated by any two specific turbines in a wind farm. The
module output (MO) signals are analysed and respective decisions are made in the decision making (DM) process. Here, turbine TZ with (Z 2 N and 1 < Z < N) represents any
turbine except T1 and TN .
H. Badihi et al. / Applied Energy 201 (2017) 284–307 289
Inconsistency
∑ A Analysis and
Isolation
Post-
Fuzzy
Processing of
∑ B Inference
Inconsistency
Mechanism
Signature
∑ C
Z NZ Z NZ
1 1
0.5 0.5
Z: Zero Z: Zero
NZ: Non-Zero NZ: Non-Zero
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Z NZ
1
0.5
Z: Zero
NZ: Non-Zero
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Rule No. If DP i;j is and P ri;j is and P gi;j is then Si;j is Pr;q ðkÞ ¼ sr;q ðkÞ xg;q ðkÞ ð14Þ
in which sr;q and xg;q are the reference generator torque and
1 Z Z Z 0
2 Z NZ Z 0.5
3 Z NZ NZ 0.5 generator angular speed for q th wind turbine in the farm,
4 Z Z NZ 1 respectively. Each particular module in the FDD system shown in
5 NZ Z Z 1 Fig. 7 has its relevant inputs and outputs. For example, Fig. 8 shows
6 NZ Z NZ 1
the relevant inputs and outputs for the example module Mi;j that
7 NZ NZ Z 1
8 NZ NZ NZ 1 monitors the consistency of the powers generated by turbine Ti
and turbine Tj in a farm (i–j). In this figure, the inputs include the
turbine. In Fig. 7, Mes is a vector of performance data including reference powers and generated powers corresponding to turbines
measured variables and control commands/references in wind tur- Ti and Tj that the module is monitoring them for power consistency,
bines. However, the FDD system only needs the sets of reference while the module output MOi;j includes inconsistency detection
powers and generated powers defined in (12) and (13), (IDi;j ) and inconsistency magnitude (IM i;j ) signals. The inconsistency
respectively. detection signal and inconsistency magnitude signal indicate,
respectively, the occurrence and magnitude of any possible
P r ðkÞ ¼ ½Pr;q ðkÞ; q ¼ 1; 2; . . . ; N ð12Þ inconsistency in the generated powers by turbines Ti and Tj . More
precisely, regardless of details of the monitoring process conducted
P g ðkÞ ¼ ½Pg;q ðkÞ; q ¼ 1; 2; . . . ; N ð13Þ in the module that are presented in two next subsections, the
Absolute Inconsistency
1
Information [-]
0.5 0.5
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
(a) (b)
Fig. 13. Post-processing of inconsistency signature in an example module: (a) inconsistency signature and (b) absolute inconsistency information.
292 H. Badihi et al. / Applied Energy 201 (2017) 284–307
module output MOi;j can attain the values described in Table 2 (s). In particular, Fig. 7 shows the distribution and contribution
under two possibilities for the consistency of power generation by of MO signals in the decision making process in each particular
turbines. DM block. As already mentioned, each DM block corresponds to
As it is observed in Table 2, when power generation is inconsis- an individual turbine and is exclusively used for analysing the
tent between Ti and Tj , the inconsistency detection and inconsis- MO signals related to that specific turbine, and then makes an
tency magnitude signals indicate the faulty turbine and the appropriate decision about the real-time status of the turbine. Such
absolute value of estimated power loss jDPi;j j, respectively. Con- a decision is simply made based on the logic which states that as
versely, when power generation is consistent between Ti and Tj , long as at least one of the relevant modules detects a turbine as
both inconsistency detection and inconsistency magnitude signals a faulty turbine, then that turbine is faulty and its power loss
are zero. This does not necessarily mean that both turbines are due to the fault (i.e., Pb q in (10)) is conservatively the maximum
healthy. In fact, when power generation is consistent between power loss estimated by the modules. In connection with Fig. 7
two turbines Ti and Tj , they can be either both healthy or both and mentioned information about Fig. 8 and Table 2, the men-
faulty. Therefore, as it is shown in Fig. 7, the DM blocks are tioned logic of decision making is formulated in (15)–(17) for T1 ,
required to further analyse the output results of modules all TZ and TN , respectively. Note that turbine TZ with (Z 2 N and
together in real-time, and consequently determine faulty turbine 1 < Z < N) represents any turbine except T1 and TN .
Table 4
^ g ðkÞ.
^ðkÞ ¼ P
Configuration properties of T-S fuzzy model (SISO model). Note that uðkÞ ¼ Pri;j ðkÞ, and y i;j
Consequent part Linear equation form in ith rule – y ^ðk 1Þ þ ai;2 uðk 1Þ þ bi
^i ðkÞ ¼ ai;1 y
PL
Defuzzification method – li ðUÞy^i
^ ¼ Pi¼1
y L
i¼1
li ðUÞ
25 25
Vnac,1 [m/S]
Vnac,6 [m/S]
20 20
15 15
10 10
5 5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
25 25
Vnac,7 [m/S]
Vnac,2 [m/S]
20 20
15 15
10 10
5 5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
25 25
Vnac,8 [m/S]
Vnac,3 [m/S]
20 20
15 15
10 10
5 5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
25 25
Vnac,4 [m/S]
Vnac,9 [m/S]
20 20
15 15
10 10
5 5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
25 25
Vnac,5 [m/S]
Vnac,10 [m/S]
20 20
15 15
10 10
5 5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
Fig. 17. Nacelle wind speeds for turbines of the wind farm shown in Fig. 1. (Note: Vnac, i denotes nacelle wind speed for turbine Ti).
1 1
0.5 0.5
0 0
-1 -0.5 0 0.5 1 1.5 -1 -0.5 0 0.5 1 1.5
6 6
x 10 x 10
(a) (b)
^ðk 1Þ, and (b) uðk 1Þ.
Fig. 19. Projected membership functions for: (a) y
Table 5 More precisely, when Si;j attains a value of 0 or 1 (i.e., rule 1 or rules
Estimated consequent parameters for the identified T-S fuzzy model with the 4–8 in Table 3), the power generation is directly determined as
structure given in Table 4.
consistent or inconsistent, respectively. Conversely, when Si;j
Rule No. (i) ai;1 ai;2 bi attains a value of 0.5 (i.e., rules 2 and 3 in Table 3), the power gen-
eration cannot be directly determined as consistent or inconsis-
1 þ8:9 101 þ1:0 101 þ3:9 102
2 þ9:0 101 þ9:1 102 4:2 102
tent. In fact, rules 2 and 3 in Table 3 are related to the so-called
response time of the whole wind farm system. That is the time
the wind farm system takes to react to its given inputs. In other
words, the wind farm system takes a short period of time (i.e.,
Table 6
Cluster centers. the response time) for Pgi;j and DP i;j to react to any change in P ri;j
by the reference power signal(s) commanded from the wind farm
Rule No. (i) ^ðk 1Þ
y uðk 1Þ
control system. During such a short response time, it is not possible
1 9:8 10þ3 7:6 10þ3 to directly determine the consistency of generated powers in a
þ3
2 5:2 10 8:0 10þ3 model-free way. Therefore, as it is shown in Fig. 9, the output of
fuzzy inference mechanism (i.e., inconsistency signature Si;j ) needs
to be post-processed in order to obtain absolute inconsistency
5
x 10 information Ii;j that can be either 0 (consistent power generation)
Model Output (Local Model 1) or 1 (inconsistent power generation). The post-processing con-
15 Model Output (Local Model 2) ducted on Si;j at the time-step k is coded based on the flowchart
Process Output
shown in Fig. 12.
Output [W]
10
As shown in the flowchart in Fig. 12, to achieve a more reliable
conclusion on Ii;j and in response to any possible occurrence of
5
inconsistency in power generation, Ii;j changes from 0 to 1 when
Si;j stays at its relevant values presented in the flowchart for T in
0
consecutive sample-times (i.e., the number of repetitions satisfies
T 1 P T in ). Similarly, Ii;j changes from 1 to 0 when Si;j stays at its rel-
-5
evant values presented in the flowchart for T out consecutive
200 210 220 230 240 250
sample-times (i.e., the number of repetitions satisfies T 2 P T out ).
Time [S] The appropriate values for T in and T out are determined by the user
through considering that the greater the values of T in and T out , the
Fig. 20. The process output and the fuzzy model output (showing the contribution
of each local model) – time period [200, 250] sec. higher the required time for changing Ii;j from 0 to 1, and 1 to 0 (i.e.,
higher conservatism). As it is shown in the flowchart in Fig. 12,
when Si;j attains value of 0.5 (i.e., during short periods of response
Table 7 time of the wind farm system described by rules 2 and 3 in Table 3),
Modelling accuracy and fitting performance of Ii;j stays at its previous value obtained in the previous time-step
the fuzzy model.
(i.e., Ii;j ðkÞ ¼ Ii;j ðk 1Þ). With respect to the flowchart shown in
VAF (%) NRMSE Fig. 12, an example of inconsistency signature and its absolute
99.2 0.004 inconsistency information are shown in Fig. 13. The figure corre-
sponds to module M 1;4 for turbines T1 and T4 while considering
the fault scenario shown in Fig. 4. As it is shown in Fig. 13, Si;j only
attains value of 0.5 for very short periods of time when the inputs
functions used for the output in which each function represents a of fuzzy inference mechanism activate rules 2 and 3 in Table 3.
single-valued constant. The absolute inconsistency information Ii;j obtained from the
With respect to the shown membership functions in Figs. 10 post-processing of inconsistency signature Si;j only determines
and 11, the complete set of rules used in the fuzzy inference mech- the presence of consistency or inconsistency in the generated pow-
anism is presented in Table 3. The rules are formulated based on an ers by turbines Ti and Tj . However, in the case of inconsistency in
expert’s knowledge obtained through observing and understand- the generated powers, Ii;j does not provide any information about
ing the relation between the consistency of generated powers isolation of faulty turbine and magnitude of fault (power loss).
and variations in the inputs. Each rule corresponds to a particular Therefore, it is required to further analyse Ii;j through inconsistency
condition. In total, three different possibilities are considered: (1) analysis and isolation block as shown in Fig. 9. The functionality of
consistent power generation represented by Si;j ¼ 0, (2) inconsis- this block is very simple in comparison with the already mentioned
tent power generation represented by Si;j ¼ 1, and (3) either consis- post-processing block. In fact, in the presence of any inconsistency
tent or inconsistent power generation represented by Si;j ¼ 0:5. in the generated powers (i.e., Ii;j ¼ 1), the inconsistency analysis
296 H. Badihi et al. / Applied Energy 201 (2017) 284–307
Power Loss
Power Loss
Estimated Fault Magnitude T1 Estimated Fault Magnitude T6
0.4 0.4
[MW]
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
Fault Indicator T1 Fault Indicator T6
1.0 1.0
[-]
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
Estimated Fault Magnitude T2 Estimated Fault Magnitude T7
[MW]
0.4 0.4
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
Fault Indicator T2 Fault Indicator T7
1.0 1.0
[-]
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
Estimated Fault Magnitude T1 Estimated Fault Magnitude T6
0.4 0.4
[MW]
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
Fault Indicator T1 Fault Indicator T6
1.0 1.0
[-]
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
Estimated Fault Magnitude T2 Estimated Fault Magnitude T7
0.4
[MW]
0.4
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Fault Indicator
Fault Indicator
[-]
[-]
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
[MW]
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Fault Indicator
Fault Indicator
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
Power Loss
Power Loss
0.2 0.2
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Table 8 the faulty turbine and obtain its magnitude of the fault (power
Time of fault detection (or detection time) for each FDD system (in loss), Ii;j obtained from the post-processing is further analysed by
seconds).
the shown inconsistency analysis and isolation block in Fig. 14.
FDD system type Detection time (s) The functionality of this block is completely similar to what was
Model-free 2.2125 already mentioned in the previous subsection for the module
Model-based 1.5000 designed using the model-free algorithm. It is also worth remind-
ing that the module M i;j proposed here provides relevant informa-
tion about detection and diagnosis of the faulty turbine only when
power generation is inconsistent. However, as already mentioned,
and isolation block computes the running mean of DPi;j in real-time when power generation is consistent between turbines Ti and Tj ,
(see Fig. 9). Then, the block determines (isolates) the faulty turbine they are either both healthy or both faulty. Therefore, as it was
based on the sign of the computed mean. That is, if the mean already shown in Fig. 7, and formulated in (15)–(17), the DM
attains positive values, Tj is faulty while if it attains negative val- blocks are required to analyse the output results of modules all
ues, it means that Ti is faulty. Moreover, the absolute magnitude together in real-time, and to consequently detect and diagnose
of this running mean jDP i;j j is used as the estimated power loss faulty turbine(s).
in the faulty turbine (see Table 2). Obviously, the performance and effectiveness of the model-
It should be noted that the module M i;j designed here only based module Mi;j proposed here depend highly on the nominal
determines whether powers generated from turbines Ti and Tj dynamic model of the system (see Fig. 14) that, as already men-
tioned, is developed using the data-driven modelling approach
are consistent or inconsistent. When power generation is inconsis-
based on FMI technique. The FMI technique generates multiple
tent, the module provides relevant information about detection
models as a collection of fuzzy if-then rules. In detail, the nominal
and diagnosis of the faulty turbine. However, as previously men-
dynamic model can be represented by a Takagi-Sugeno (T-S) type
tioned at the beginning of Section 5, when power generation is
dynamic fuzzy model identified from input/output of the system’s
consistent between turbines Ti and Tj , they are either both healthy
measurements obtained from the simulation of the system under
or both faulty. More precisely, if the fault happens simultaneously
normal (fault-free) operation. Such a method of modelling is par-
in both turbines Ti and Tj , the module M i;j still shows consistent
ticularly suitable for wind turbines that have relatively complex
power generation. For example, this is shown in Fig. 13 for module
structures comprising flexible mechanical systems immersed in a
M 1;4 while both turbines T1 and T4 are faulty during [225, 300] sec-
fully-stochastic wind field. In practice, it is often demanding, or
onds. To address this issue, as it was previously shown in Fig. 7,
even impossible to derive a single nonlinear model for wind energy
and formulated in (15)–(17), the DM blocks are designed and
systems over their entire operational envelope. In the following,
implemented in the FDD system to further analyse the output
the utilized FMI technique is described briefly.
results of modules all together in real-time, and to consequently
Without loss of generality, a single-input single-output (SISO)
detect and diagnose faulty turbine(s).
nonlinear system can be expressed by:
yðkÞ ¼ f ðUðk 1ÞÞ þ e ð19Þ
5.2. Model-based monitoring of power consistency
where Uðk 1Þ as defined in (20) is an information data vector
From R modules used in the FDD system shown in Fig. 7, con- including the past model inputs u and outputs y, k is the discrete-
sider an example module M i;j that monitors the consistency of time-step, fm; ng 2 Z denote the model order that are defined by
the powers generated by turbines Ti and Tj in a farm (i–j). This sub- the user, and e denotes the modelling error.
2 3
section addresses the design of such a module using a model-based yðk 1Þ
algorithm based on an FMI technique. A block diagram illustrating 6 .. 7
6 . 7
the overall structure of the module designed using the above- 6 7
6 7
mentioned model-based algorithm is shown in Fig. 14. 6 yðk mÞ 7
As shown in Fig. 14, the module employs a nominal dynamic Uðk 1Þ ¼ 6 7
6 uðk 1Þ 7 ð20Þ
6 7
model of the system that is developed using a data-driven mod- 6 7
6 .. 7
elling approach based on FMI technique. The model estimates the 4 . 5
nominal relative performance between turbines Ti and Tj in a farm. uðk nÞ
In fact, powers generated from turbines Ti and Tj are consistent as
^ g and P g are (ideally) equal. However, they cannot be A T-S type fuzzy model can approximate the unknown function
long as P i;j i;j
f ð:Þ in (19) using L rules (local models) as follows [25]:
equal in practice due to the presence of noises and uncertainties
inherent in real data. Therefore, a post-processing including a Rulei : ^i ðkÞ ¼ F i ðUðk 1ÞÞ;
IfUðk 1Þis Ai theny i ¼ 1; 2; . . . ; L
threshold testing is necessary to obtain absolute inconsistency ð21Þ
information Ii;j that can be either 0 (consistent power generation)
or 1 (inconsistent power generation). The post-processing at the where Ai are the antecedent fuzzy sets of the ith rule. A simple but
time-step k is coded based on the flowchart shown in Fig. 15. Sim- efficient linear form for F i ð:Þ is defined by [25]:
ilar to what was mentioned in the previous subsection, to achieve a ^i ðkÞ ¼ ai Uðk 1Þ þ bi
F i ð:Þ : y ð22Þ
more reliable conclusion on Ii;j , in response to any possible occur-
rence of inconsistency in power generation, Ii;j changes from 0 to where ai and bi denote the parameter vector and scalar offset of the
1 when residual r stays outside of its threshold for T in consecutive ith rule, respectively.
sample-times (i.e., the number of repetitions satisfies T 1 P T in ). The following fuzzy fusion over all rule outputs provides the
Similarly, Ii;j changes from 1 to 0 when residual r stays within its dynamics of a nonlinear system [25]:
!, !
threshold for T out consecutive sample-times (i.e., the number of X
L X
L
Fault-Free Operation
Faulty Operation with AFTCC
Faulty Operation without AFTCC
5 5
4 4
3 3
4
2 2 4
1 1
4 4
3 4 3
3 4
2 2 2
1
1 0 1
500 600 700 800 900 1000 150 200 250 300
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
4 4
3 3
2 2
4
4
1 1 3
400 500 600 700 800 900 1000 200 400 600 800 1000
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
(e) (f)
5
Generator Power [MW]
4
2
1 3
Fault-Free Operation
Faulty Operation with AFTCC
Faulty Operation without AFTCC
5 5
Generator Power [MW]
3 3
4
2 2 4
1 1
3 3
4
2 2
4
1 1
500 600 700 800 900 1000 150 200 250 300
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
4 4
3 3
2 2
4
4
1 1 3
400 500 600 700 800 900 1000 200 400 600 800 1000
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
(e) (f)
5
Generator Power [MW]
4
2
1 3
Table 9 Table 10
Quantitative comparison of generator power responses for AFTCC schemes during the Quantitative comparison of generator power responses for AFTCC schemes during the
specified fault periods with 3% power loss. specified fault period with 30% power loss.
Faulty turbine Fault period [S] NRMSE [–] Faulty turbine Fault period [S] NRMSE [–]
Model-based AFTCC Model-free AFTCC Model-based AFTCC Model-free AFTCC
T1 [225, 400] 0.0031 0.0037 T10 [700, 1000] 0.0266 0.0304
T2 [800, 1000] 0.0030 0.0035
T3 [450, 1000] 0.0018 0.0021
T4 [125, 300] 0.0032 0.0038
T7 [350, 1000] 0.0015 0.0017 fault-free simulation with the baseline wind farm controller in (5)
T8 [100, 1000] 0.0023 0.0033 is used as a frame of reference to evaluate the overall performance,
T10 [700, 1000] 0.0025 0.0030 effectiveness, and fault-tolerance properties of the proposed FDD
and FTC schemes against the considered decreased power genera-
^ is the aggregated output of the fuzzy model, and li denotes tion faults that are already described in Fig. 4. In this regard, with
where y
respect to the considerations given in Section 3, the generator
the degree of fulfilment of ith rule in (21).
power response is chosen as a well-suited performance index to
To develop the above-mentioned T-S model for a real system, it
illustrate the performance of each turbine in the wind farm.
is necessary to perform the nonlinear system identification that is
the process of identifying the structure of fuzzy model, and then
estimating the parameters in the model. The detailed structure of
6.1. Identification and validation of the fuzzy dynamic model
the fuzzy model developed for application in the module in
Fig. 14 is presented in Table 4. It is worth mentioning that in order
As already mentioned, the nominal dynamic model used in the
to identify the T-S model, the well-established Gustafson-Kessel
modules of the model-based FDD system is designed and devel-
clustering algorithm [26] is used here.
oped using the FMI technique described in Section 5.2. In more
details, to train and evaluate a model, a set of 80,000 measured
6. Simulation results and discussion data for each of the inputs and outputs were used. The data
was obtained with a sampling rate of 80 Hz from the fault-free
This section presents and discusses different simulation results simulation of the wind farm with its baseline control system. It
for the investigation of the performance of proposed schemes is worth mentioning that each set of the data was split into equal
under both fault-free and faulty conditions. The simulations in this halves; one half for training and the other one for validation. The
work have been carried out in MATLAB/Simulink environment structure for the fuzzy model is already determined in Table 4. In
using the nonlinear offshore wind farm benchmark model pre- connection to the mentioned table and the performed nonlinear
sented in Section 2. The general outline of the simulation platform system identification process, the projected membership func-
is shown in Fig. 16. tions (two membership functions per input) are shown in
In more details, an offshore wind farm including ten 5 MW- Fig. 19. The consequent parameters of the model as well as the
turbines (N ¼ 10) is created with the layout as shown in Fig. 1. A cluster centers are also presented in Tables 5 and 6, respectively.
realistic wind field with a mean speed of 15 m/s and a turbulence In addition, for example, Fig. 20 shows the output of a fuzzy
intensity of 10% is simulated over 1000 s of run time. The obtained model during a fault-free operation of the wind farm including
wind speed profiles for each of the ten turbines in the farm are the contribution of each local model (rule) in the developed
shown in Fig. 17. Moreover, a typical electrical grid load is applied model over a period of 50 s of simulation time. In this figure,
to the wind farm over the simulation time. Fig. 18 shows the the noise-free measurement of process output is used for compar-
applied time-varying grid load as well as the relevant total gener- ison with the model output.
ated active power response obtained from the wind farm under To assess the quality of the developed fuzzy model in terms of
fault-free conditions. modelling accuracy and fitting performance, the normalized root-
In the following subsections, various simulations and numerical mean-squared error (NRMSE) and the variance accounted for (VAF)
results for the proposed schemes are presented and discussed. The index are used, respectively. The NRMSE is defined as:
4 4
3 3
2 2
1 1
0 0
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
(a) (b)
Fig. 25. Generator power response under fault-free and faulty conditions (30% power loss) in T10: (a) model-free AFTCC, and (b) model-based AFTCC.
302 H. Badihi et al. / Applied Energy 201 (2017) 284–307
x 10-3 x 10-3
4 4
Fault-Free Operation Fault-Free Operation
[Rad/S]
[Rad/S]
0 0
-2 -2
-4 -4
0 200 400 600 800 1000 0 200 400 600 800 1000
Time [S] Time [S]
x 10-3 x 10-3
2 2
1 1
[Rad/S]
[Rad/S]
0 0
-1 -1
-2 -2
220 225 230 235 220 225 230 235
Time [S] Time [S]
Shaft Torsion Angle
Shaft Torsion Angle
x 10 -3 x 10-3
2 2
1 1
[Rad/S]
[Rad/S]
0 0
-1 -1
-2 -2
390 395 400 405 410 390 395 400 405 410
Time [S] Time [S]
(a) (b)
Fig. 26. Drivetrain torsion rate results for wind turbine T1 in the farm during the: (a) model-free AFTCC, and (b) model-based AFTCC.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
XN
1
ðy y ^k Þ2 indicators/signals and estimated fault magnitudes that can be used
N k¼1 k
NRMSE ¼ ð24Þ for ASC and accommodation of the faults in faulty turbines.
ymax ymin The time of fault detection (i.e., the time required for detecting a
fault after its occurrence) for each FDD system is presented in
where yk and y^k are the kth true output of the system (process), and
Table 8. The time of fault detection reflected in this table is directly
the kth estimated output of the model, respectively. The denomina-
related to the T in consecutive sample-times considered in the
tor in (24) represents the range of the true output defined as the
design process of each FDD system. As can be seen from Table 8,
maximum value ymax minus the minimum value ymin . The percentile
the model-based FDD system requires lower time for detection of
VAF is computed by:
faults. The lower the detection time, the better the fault detection
cov ðyk y
^k Þ has scored.
VAF ¼ 1 100% ð25Þ With respect to the identification of faults, both model-free and
cov ðyk Þ
model-based FDD systems provide almost similar estimates for
where cov denotes the covariance of the respective vector. Table 7 fault magnitudes in each particular faulty turbine. The fault magni-
presents the obtained results. From the table, it is obvious that tude itself represents the estimated amount of power loss that
the fuzzy model is considerably accurate for approximating the pro- actually happens due to the occurrence of fault in a turbine. As it
cess under diagnosis. is shown in Figs. 21 and 22, the fault magnitudes are estimated
exactly during the time that the fault detection indicators/signals
are generated. These estimates are accurate enough to be used
6.2. Performance of FDD system for the fault accommodation process. This fact is quantitatively
shown in the next subsection while the obtained estimates are
As it was previously presented in Section 5, the FDD system can used directly in (10) for signal correction and fault accommodation
be developed based on either model-free or model-based monitor- in the framework of integrated FDD and FTC approach.
ing of power consistency in the wind farm. Both model-free FDD
system and model-based FDD system are investigated under both
fault-free and faulty conditions. 6.3. Performance of AFTCC schemes based on integrated FDD and FTC
As it is shown by the FDD results in Figs. 21 and 22, the pro- approach
posed model-free and model-based FDD systems both are able to
detect and diagnose all the considered faults based on the scenario As already mentioned, the estimated fault magnitudes act upon
defined in Fig. 4. The faults occur in turbines T1, T2, T3, T4, T7, T8, the reference torque control signal (see (10)) to accommodate the
and T10. The FDD results are shown in terms of both fault detection fault effects in each faulty turbine in the farm. Based on the
H. Badihi et al. / Applied Energy 201 (2017) 284–307 303
Fault-Free Operation
Faulty Operation with AFTCC
8
x 10 8
Tower Moment
Tower Moment
x 10
1 1
0.5 0.5
[N.m]
[N.m]
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(a) (b)
8 8
x 10 x 10
Tower Moment
Tower Moment
1 1
0.5 0.5
[N.m]
[N.m]
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(c) (d)
8 8
x 10 x 10
Tower Moment
Tower Moment
1 1
0.5 0.5
[N.m]
[N.m]
0 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(e) (f)
7
x 10
Tower Moment
0
[N.m]
-5
-10
0 100 200 300 400 500 600 700 800 900 1000
Time [S]
(g)
Fig. 27. Tower bending moment results model-free AFTCC: (a) T1, (b) T2, (c) T3, (d) T4, (e) T7, (f) T8, and (g) T10.
integrated FDD and FTC approach presented in Section 4, two only activated during the periods of fault activity, the nominal
AFTCC schemes are designed and developed using the model-free performance of wind turbines will be unaffected under fault-free
and model-based FDD systems in Section 5. In the following figures conditions.
and tables, the AFTCC scheme using the model-free FDD system is To further investigate the performance of the proposed schemes
denoted by ‘‘model-free AFTCC” scheme, and the AFTCC scheme in a comparative perspective, Table 9 presents a precise quantita-
using the model-based FDD system is denoted by ‘‘model-based tive comparison between both schemes in terms of NRMSE during
AFTCC” scheme. the periods of fault activity in the faulty wind turbines. Note that
Based on the FDD results presented in the preceding subsection, the lower the NRMSE, the better the fault-tolerance has scored.
the fault accommodation is performed in the faulty turbines that Based on the results presented in Table 9, the model-based AFTCC
include T1, T2, T3, T4, T7, T8, and T10. Figs. 23 and 24 show the gen- scheme provides a more efficient accommodation of faults com-
erator power responses in the mentioned turbines under fault-free pared to the model-free AFTCC scheme. This is due to the fact that
conditions, faulty conditions with AFTCC schemes, and faulty con- the performance of each scheme was highly dependent on the
ditions without AFTCC schemes. speed and accuracy of its FDD system. As already discussed in
Fig. 23 represents the performance of the model-free AFTCC the previous subsection, the model-based FDD system demon-
scheme, while Fig. 24 represents the performance of the model- strates better performance compared to the model-free FDD sys-
based AFTCC scheme. Both figures illustrate that the accommoda- tem that does not employ any dynamic model of the system.
tion of faults is successfully performed in the faulty turbines and However, this superior performance of the model-based FDD sys-
during all periods of fault activity. In other words, the performance tem may be subject to modelling uncertainties, while the model-
of the system under faulty conditions with AFTCC schemes remains free system does not deal with any difficulties related to system
as close as possible to the performance of the system under fault- modelling.
free conditions (i.e., the system’s nominal performance). Moreover, In addition to the presented results, Fig. 25 together with the
as foreseen, since the control reconfiguration (signal correction) is numerical results presented in Table 10 show the fault-tolerance
304 H. Badihi et al. / Applied Energy 201 (2017) 284–307
Fault-Free Operation
Faulty Operation with AFTCC
8 8
x 10 x 10
Tower Moment
1 1
Tower Moment
0.5 0.5
[N.m]
0 0
[N.m]
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(a) (b)
8 8
x 10 x 10
Tower Moment
1 1
Tower Moment
0.5 0.5
[N.m]
0 0
[N.m]
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(c) (d)
8 8
x 10 x 10
Tower Moment
1 1
0.5 Tower Moment 0.5
[N.m]
0 [N.m] 0
-0.5 -0.5
-1 -1
-1.5 -1.5
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
Time [S] Time [S]
(e) (f)
7
x 10
Tower Moment
0
[N.m]
-5
-10
0 100 200 300 400 500 600 700 800 900 1000
Time [S]
(g)
Fig. 28. Tower bending moment results model-based AFTCC: (a) T1, (b) T2, (c) T3, (d) T4, (e) T7, (f) T8, and (g) T10.
Table 11
Quantitative comparison of tower bending moment results during fault periods.
Faulty turbine Fault period [s] Fault-free operation Faulty operation with AFTCC
Mean [kN m] STD [kN m] Model-based AFTCC Model-free AFTCC
Mean [kN m] STD [kN m] Mean [kN m] STD [kN m]
T1 [225, 400] 28,574 6166.2 29,453 6291.4 29,453 6294.6
T2 [800, 1000] 36,041 10,278 37,320 11,379 37,317 11,376
T3 [450, 1000] 34,570 8555.2 35,740 8989.5 35,739 9000.6
T4 [125, 300] 28,807 5243.3 29,794 7000.1 29,794 7000.9
T7 [350, 1000] 33,755 7202.2 34,824 7414.3 34,825 7412.5
T8 [100, 1000] 31,985 8473.4 33,005 8623.4 33,008 8629.5
T10 [700, 1000] 33,837 7426.5 34,977 8229.8 34,975 8227.2
capability of the proposed AFTCC schemes against a more serious 6.4. Evaluation of wind farm structural dynamics and loading results
fault with 30% power loss in turbine T10 as an example. This
demonstrates that both of the schemes can also satisfactorily While evaluating the performance and effectiveness of the pro-
accommodate larger magnitudes of the considered type of fault posed schemes, it is also necessary to evaluate the functionality of
in turbines. However, this does not eliminate the final necessity the proposed schemes in terms of wind farm structural dynamics
for maintenance of the blades before the amount of debris build and loading. With respect to the implementation of control
up and/or erosion reaches a highly serious level. schemes, the recommended rate and magnitude limiters are
H. Badihi et al. / Applied Energy 201 (2017) 284–307 305
Table 12
The results of Monte Carlo simulation studies under wind field shown in Fig. 17.
wind field, similar simulations as described in Section 6.3 are con- until the next maintenance time is reached or scheduled for faulty
ducted with the exception of fault activity periods that are parts replacement and/or investigating the root causes which
extended by an extra 1000 s of run time. Table 13 presents the might also include blade material failure. The proposed AFTCC
quantitative comparison between the proposed schemes in terms schemes in fact lessen the downtime of wind turbines and extend
of NRMSE during the mentioned periods of fault activity in the their availability in a wind farm, in order to achieve higher cost-
faulty wind turbines (the simulation plots are not shown here for efficiency in energy generation. However, it does not eliminate
the sake of brevity). the final necessity for maintenance of the blades before the
All results presented in Tables 12 and 13 confirm that both amount of debris build-up and/or erosion reaches a highly serious
model-free and model-based AFTCC schemes are not only robust level.
in the presence of modelling errors, measurement uncertainties, With respect to future research topics related to this presented
and external disturbances, but also can successfully maintain reli- work, it can be promising to extend the proposed AFTCC approach
able wind farm performance under faulty conditions. to accommodation of other frequent faults in wind farms, such as
misalignment of one or more blades originated at the time of
7. Conclusions and future works installation and/or change in the drivetrain damping due to wear
and tear. Moreover, it should be noted that in order to practically
This paper proposed a novel integrated fault detection and diag- implement and benefit from such a fault-tolerant control approach,
nosis (FDD) and fault-tolerant control (FTC) approach in a cooper- the FDD system and real-time control system reconfiguration as a
ative framework referred to as active fault-tolerant cooperative whole needs to be designed and developed along with the so-
control (AFTCC) oriented to the design and development of two called techniques related to fault-tolerant computing and fault-
AFTCC schemes for an offshore wind farm. The designed AFTCC tolerant communication networks that are followed by
schemes can address decreased power generation faults caused hardware-in-the-loop testing for complete evaluation before real
by turbine blade erosion and debris build-up on the blades over system tests.
time. Each of the designed schemes employs a FDD system to pro-
vide accurate and timely diagnosis information to be used in an Acknowledgments
appropriate automatic signal correction algorithm for accommoda-
tion of faults in the farm. The first scheme is based on a model-free This work is partially supported by the Natural Sciences and
FDD system that incorporates a rule-based threshold testing tech- Engineering Research Council of Canada (NSERC), a Seed Fund of
nique for residual evaluation. Conversely, the second scheme is Concordia University, and NSFC #61573282 and SPNSF
based on a model-based FDD system that incorporates data- #2015JZ020. The authors would like to express their appreciation
driven models developed using fuzzy modelling and identification to Drs. P.F. Odgaard, J. Stoustrup and K.E. Johnson for the wind tur-
technique. bine benchmark models in [7] and [12], the wind farm benchmark
Different simulations have been performed using a high-fidelity model used in this work from Drs. M. Soltani, T. Knudsen, and T.
offshore wind farm benchmark model in the presence of wind tur- Bak’s group [18], and Dr. S. Simani for the provided information
bulences, measurement noises, and realistic fault scenarios. More- on the relevant fuzzy modelling, identification and control tech-
over, extensive Monte Carlo simulations are performed to evaluate niques. The authors would also like to thank Dr. J. Stoustrup for
the robustness of the designed schemes with respect to modelling the continuous encouragement for authors to get into this field
errors, disturbances and measurement uncertainties. All simula- of wind turbine research. The authors would like further to extend
tion studies and numerical results clearly indicate the effectiveness their sincere thanks and appreciation to the Editor-in-Chief, Asso-
and robustness of the schemes over the entire range of tested wind ciate Editor and anonymous Reviewers for handling review to our
profiles for both the fault-free and faulty conditions. manuscript, and the constructive comments and suggestions given
Both AFTCC schemes are superior to the baseline wind farm during the review process for us to be able to significantly improve
control system designed using classical methods. The fault- the quality of the paper. Finally, the authors would like to thank
tolerance and nominal performance provided by the schemes Mr. Sean Zhang and Mr. Christopher Zhang for the proofreading
make them an efficient and practical choice for wind farms. In par- and comments that helped to improve the manuscript presenta-
ticular, the fact that the proposed schemes employ an automatic tion significantly.
signal correction algorithm for fault accommodation, and hence
do not disturb the nominal performance of a wind farm under nor-
References
mal (fault-free) operating conditions, is a remarkable feature. This
feature is particularly favorable in the case of any already designed [1] Sloth C, Esbensen T, Stoustrup J. Robust and fault-tolerant linear parameter-
wind farm control systems whose structures are well-defined, for varying control of wind turbines. Mechatronics 2011;21:645–59.
optimizing power capture in a farm while achieving fault- [2] Tabatabaeipour SM, Odgaard PF, Bak T, Stoustrup J. Fault detection of wind
turbines with uncertain parameters: a set-membership approach. Energies
tolerance capability without scarifying the optimal performance 2012;5:2424–48.
of the wind farm. However, it is worth mentioning that the perfor- [3] Simani S, Castaldi P. Active actuator fault-tolerant control of a wind turbine
mance of the proposed AFTCC schemes is highly dependent on the benchmark model. Int J Robust Nonlinear Control 2014;24:1283–303.
[4] Badihi H, Zhang YM, Hong H. Fuzzy gain-scheduled active fault-tolerant
speed and accuracy of their FDD systems. Obviously, it may be control of a wind turbine. J Franklin Inst 2014;351:3677–706.
impossible in practice to measure or obtain the instant precise val- [5] Badihi H, Zhang YM, Hong H. Wind turbine fault diagnosis and fault-tolerant
ues of faults in a system. In fact, the FDD systems only determine torque load control against actuator faults. IEEE Trans Control Syst Technol
2015;23:1351–72.
an online estimate for the magnitude of faults after their occur- [6] Odgaard P, Stoustrup J, Kinnaert M. Fault tolerant control of wind turbines - a
rence while accepting a detection time delay. benchmark model. In: The 7th IFAC symposium on fault detection, supervision
It is worth mentioning that the considered fault of blade erosion and safety of technical processes, Barcelona, Spain. p. 155–60.
[7] Odgaard PF, Johnson KE. Wind turbine fault diagnosis and fault tolerant control
and debris build-up on the blades actually happens over the oper-
- an enhanced benchmark challenge. In: Proc. of American Control Conference,
ation time of a wind turbine in a wind farm. The AFTCC schemes Washington, DC, USA.
proposed in this paper are basically developed for timely detection [8] Badihi H, Zhang YM, Hong H. A review on application of monitoring, diagnosis,
and diagnosis and accommodation of such a fault in a wind farm, and fault-tolerant control to wind turbines. In: Proc. of International
Conference on Control and Fault-Tolerant Systems, Nice, France.
and finally, extending the operation time (availability) of any faulty [9] Kusiak A, Verma A. A data-driven approach for monitoring blade pitch faults in
wind turbine with an acceptable (probably degraded) performance wind turbines. IEEE Trans Sustain Energy 2011;2:87–96.
H. Badihi et al. / Applied Energy 201 (2017) 284–307 307
[10] Kusiak A, Verma A. A data-mining approach to monitoring wind turbines. IEEE [19] Badihi H, Zhang YM, Hong H. Active fault tolerant control in a wind farm with
Trans Sustain Energy 2012;3:150–7. decreased power generation due to blade erosion/debris build-up. IFAC-
[11] Marvuglia A, Messineo A. Monitoring of wind farms’ power curves using PapersOnLine 2015;48:1369–74.
machine learning techniques. Appl Energy 2012;98:574–83. [20] Zhang YM, Jiang J. Bibliographical review on reconfigurable fault-tolerant
[12] Odgaard PF, Stoustrup J. Fault tolerant wind farm control — a benchmark control systems. Ann Rev Contr 2008;32:229–52.
model. In: Proc. of IEEE international conference on control applications, [21] Soltani M, Knudsen T, Bak T. Modeling and simulation of offshore wind farms
Hyderabad, India. p. 412–7. for farm level control. In: European offshore wind conference and exhibition,
[13] Simani S, Farsoni S, Castaldi P. Residual generator fuzzy identification for wind Stockholm, Sweden; 2009.
farm fault diagnosis. In: Proc. of the 19th world congress of the international [22] Jonkman J, Butterfield S, Musial W, Scott G. Definition of a 5 MW reference
federation of automatic control (IFAC), Cape Town, South Africa. wind turbine for offshore system development. Colorado, USA: National
[14] Borcehrsen AB, Larsen JA, Stoustrup J. Fault detection and load distribution for Renewable Energy Laboratory; 2009.
the wind farm challenge. In: Proc. of the 19th world congress of the [23] Pao LY, Johnson KE. Control of wind turbines: approaches, challenges, and
international federation of automatic control (IFAC), Cape Town, South Africa. recent developments. IEEE Control Syst Mag April 2011;31:44–62.
[15] Simani S, Farsoni S, Castaldi P. Fault-tolerant control of an offshore wind farm [24] Fingersh L, Carlin P. Results from the N R E L Variable-Speed Test Bed. In: Proc.
via fuzzy modelling and identification. IFAC-PapersOnLine 2015;48:1345–50. of the 17th ASME Wind Energy Symposium, Reno, USA. p. 233–7.
[16] Odgaaard PF, Shafiei SE. Evaluation of wind farm controller based fault [25] Babuska R. Fuzzy Modeling for Control. Springer: Kluwer Academic Publishers;
detection and isolation. IFAC-PapersOnLine 2015;48:1084–9. 1998.
[17] Duviella E, Serir L, Sayed-Mouchaweh M. An evolving classification approach [26] Gustafson DE, Kessel WC. Fuzzy Clustering with a Fuzzy Covariance Matrix. In:
for fault diagnosis and prognosis of a wind farm. In: Proc. of the 2nd IEEE Conference on Decision and Control, San Diego, USA. p. 761–6.
conference on control fault-tolerant systems, Nice, France. p. 377–82.
[18] Blesa J, Jiménez P, Rotondo D, Nejjari F, Puig V. An interval NLPV parity
equations approach for fault detection and isolation of a wind farm. IEEE Trans
Industr Electron 2015;62:3794–805.