Sie sind auf Seite 1von 8

Received: 27 November 2018 Revised: 9 January 2019 Accepted: 9 January 2019

DOI: 10.1002/er.4382

TECHNICAL NOTE

Wind turbine generation performance monitoring with


Jaya algorithm

Rui Jin1 | Long Wang1,2 | Chao Huang1 | Shancheng Jiang3

1
School of Computer and Communication
Summary
Engineering, University of Science and
Technology Beijing (USTB), Beijing Wind turbine (WT) power curves effectively reflect the generation performance
100083, China of WTs and depict the relationship between the wind speed and the WT power
2
School of Data Science, City University of output. This paper aims at developing an effective method for learning the
Hong Kong, Tat Chee Avenue, Kowloon,
Hong Kong intrinsic representations of WT power curves, which are robust to external envi-
3
Department of Industrial and Systems ronmental changes. Based on the obtained representations, WT generation per-
Engineering, The Hong Kong Polytechnic formance is monitored. In the proposed approach, data of the supervisory
University, Hung Hom, Kowloon, Hong
Kong
control and data acquisition (SCADA) system is employed to derive the repre-
sentations. Parametric models of WT power curves are developed using the
Correspondence two‐parameter and four‐parameter logic models. The parameters of these model
Long Wang, School of Computer and
Communication Engineering, University
are identified via Jaya algorithm. To detect the changes of WT power curve
of Science and Technology Beijing model parameters over different time, multivariate control charts are employed.
(USTB), Beijing 100083, China. The effectiveness of the proposed WT generation performance monitoring
Email: long.wang@ieee.org Shancheng
Jiang, Department of Industrial and approach is validated based on SCADA data collected from real commercial
Systems Engineering, The Hong Kong WTs.
Polytechnic University, Hung Hom,
Kowloon, Hong Kong. KEYWORDS
Email: shancheng.jiang@polyu.edu.hk
Jaya algorithm, multivariate approach, performance monitoring, power curve

Funding information
Guangdong Provincial Key Laboratory of
New and Renewable Research and Devel-
opment, Grant/Award Number:
Y807s61001; University of Science and
Technology Beijing—National Taipei
University of Technology Joint Research
Program, Grant/Award Number:
TW2018008; Fundamental Research
Funds for the Central Universities, Grant/
Award Number: 06500078; National Nat-
ural Science Foundation of China, Grant/
Award Number: 71473155

1 | INTRODUCTION plays a significant role in clean power generation, providing


a large quantity of electricity without generating carbon
During recent years, limited fossil‐fuel resources, global dioxide emissions. However, the uncertainty of wind
warming, and other environmental disasters draw a public turbine (WT) conditions, such as the significant WT down-
concern. Governments around the world have widely time, may influence the operations of the power grid.2,3
released targets of pushing the adoption of renewable In wind energy domain, various research on WT
resources.1 Among various renewable sources, wind energy condition monitoring (CM) and fault diagnosis3,4 has

Int J Energy Res. 2019;1–8. wileyonlinelibrary.com/journal/er © 2019 John Wiley & Sons, Ltd. 1
2 TECHNICAL NOTE

been carried out to reduce the non‐working time and


increase the power output. As supervisory control and Novelty Statement
data acquisition (SCADA) systems have been mounted
The four‐parameter logistic model is employed
on most commercial WTs, no more hardware investment
for WT generation performance monitoring to
is required when developing a SCADA data‐based CM sys-
improve the monitoring performance in this
tem.5 The value of SCADA data varies within a large inter-
paper. The Jaya algorithm is applied to estimate
val for different working conditions, and thus it is difficult
the logistic model parameters without using any
to detect early anomalies without using appropriate data
prior knowledge. The proposed approach is
analysis tools. To address this issue, various parametric
compared with the recently published method
and nonparametric models were developed to depict
to prove its superiority.
the WT power curves (WTPCs) based on SCADA data,
because the WTPC essentially reflects the generation per-
formance of WTs.6-8 Because parametric models can offer anomalies can be improved. In the proposed monitoring
decent modeling performance and are easy to be deter- approach, the obtained model parameters are monitored
mined, they are widely utilized in the wind energy
by applying the multivariate control chart. Figure 1 illus-
industry. Long et al9 developed linear and nonlinear para- trates the online monitoring procedures by applying the
metric models to fit the SCADA data and obtain the proposed approach.
WTPCs. Based on the results of WTPCs, multivariate and
As described in Figure 1, the proposed framework
residual approaches were applied to detect anomalies. includes five steps:
Lydia et al6 applied advanced optimization algorithms
to estimate parameters of parametric models with four
Step 1: collect SCADA data from the wind farm.
and five‐parameter logistic expressions. Because WTPC Step 2: filter the data obtained from each WT.
models are determined by the model parameters, parame- Step 3: divide the filtered data into sub‐datasets with 1‐
ter estimation methods have been widely studied. As day interval and generate into profiles.
shown in previous studies,10-12 the least‐square methods Step 4: load profiles into CMS and calculate the param-
were usually employed to estimate parameters of WTPCs. eter vector.
However, prior knowledge about these estimated parame- Step 5: apply the multivariate approach to generate
ters are required, and improper parameter initializations
control charts and obtain outliers.
may lead to non‐optimal solutions. To address this issue,
parameter estimation approaches using the PSO algorithm The main contributions of this paper are listed as follows:
were discussed.13,14 Although the PSO algorithm can
achieve better estimation performance without using any 1. The four‐parameter logistic model is employed for
prior knowledge, some algorithm‐specific parameters are WT generation performance monitoring to improve
introduced by using the PSO algorithm, and tuning these the monitoring performance.
parameters needs more computational cost. 2. The Jaya algorithm is applied to estimate the logistic
In this paper, we apply a four‐parameter logistic model model parameters without using any prior knowledge.
to model WTPCs. Meanwhile, Jaya algorithm is utilized to 3. The proposed approach is compared with the recently
extract wind power curve parameters. Because Jaya
published method to prove its superiority.
algorithm15 is a meta‐heuristic algorithm without any
algorithm‐specific parameters, it has the capacity of address-
ing the two main existing issues in previous parameter esti- 2 | DATA SUMMARY
mation studies. Besides, the employed four‐parameter
logistic model can better describe the WTPCs compared The 10‐minute SCADA data are used in this study, which
with the linear model9 so that the detection rate of is collected from 32 WTs of 1.6‐MW class at a commercial

FIGURE 1 The online monitoring procedures using the proposed approach [Colour figure can be viewed at wileyonlinelibrary.com]
TECHNICAL NOTE 3

TABLE 1 Parameters recorded by the SCADA system where m denotes the gradient, c depicts the WTPC intercept,
and Pmax denotes the largest WT power output. The model
Name 1# Blade Angle
is fitted by the least square method through minimizing
Time 1# set pitch angle the summed square of residuals, presented as Equation 2:
Grid side frequency 2# blade angle N
Generator active power 2# set pitch angle D ¼ ∑ ðPe ðvi Þ−Pat ðiÞÞ2 (2)
i¼1
Generator speed 3# blade angle
where Pe(vi) is the fitted response value, and Pat(i) is the
Nacelle temperature 3# set pitch angle
actual power output. Based on Equations 1 and 2, the char-
Wind direction Temperature of gearbox lubricant oil
acteristics of WPTC could be depicted as (mt, ct), where t is
Wind speed The rate of available utilization the sequence number of samples.

wind farm located in Jiangsu Province, China, during the


4.2 | Four‐parameter logistic model
period of 1 April 2014 to 30 July 2014. The parameters
recorded by SCADA system are listed in Table 1.
Because the shape of WTPC is similar to that of the
cumulative logistic distribution function, the WTPC can
3 | DATA FILTERING A ND be described as a logistic function using four parame-
PARTITIONING ters,6,16 shown as Equation 3:
1 þ me−τ
v

The speed of wind and power output of WTs are considered P ¼ f ðv; θÞ ¼ a v: (3)
1 þ ne−τ
to develop WTPC models in this study, and thus the used
dataset can be depicted as X = {(P1, v1), (P2, v2), … , (PN, In Equation 3, θ = (a, m, n, τ) denoting the parameter set
vN)}, where Pi is the WT power output of the i‐th data sam- of the logistic model and θ determinates shapes of
ple, vi is the wind speed of the i‐th sample, and N denotes WTPCs. In this model, D also serves as the objective func-
the number of data samples. The data consist of totally 119 − τi
v

samples with 1‐day interval (each sample is a pile of dataset tion to be minimized, and Pe ðvi Þ ¼ a1 þ mev
− i
is the esti-
1þne τ

stored in a file, and the data in the period of 12 April to 13 mated power output. The constraints are shown as
April are lost). Different from the previous study using the follows in Equation 4:
2‐day interval,9 the 1‐day interval can capture more WTPC 
amin < a < amax ; mmin < m < mmax
variations during the considered monitoring time period. : (4)
In addition, the cut‐in, rated, and cut‐out speeds (ie, nmin < n < nmax ; τ min < τ < τ max
vci, vr, and vco) of involved WTs are 3.0, 10.0, and
21.0 m/s separately. Data points with wind speed out of In order to obtain the optimally fitted parameter vector,
the range [cut‐in speed, cut‐out speed] are removed. Jaya algorithm is introduced to obtain the solution of this
optimization problem. Compared with other popular heu-
ristic search algorithms, including PSO algorithm and
4 | W I N D P O W E R CU R V E Genetic algorithm, Jaya algorithm provides the supreme
MODELING performance in both constraint and unconstraint bench-
mark problems,15,17-19 and objective functions of these
Two parametric models for WTPC modeling, the linear- problems demonstrate different characteristics. Algo-
ized segmented model and four‐parameter logistic model, rithm 1 illustrates the mechanism of Jaya algorithm.
have been discussed in detail in this section. As presented in Algorithm 1, the solution, parameter
vector θ, is updated according to Equation 5,
  
4.1 | Linearized segmented model θ′j;k;i ¼ θj;k;i þ r 1; j;i θj;best;i − θj;k;i 
  
− r 2; j;i θj;worst;i − θj;k;i  (5)
The dataset is divided into three subsets according to the
conventional WTPC, and the linearized segmented model
where θj, k, i denotes the value of the jth element in kth
is shown as follows9:
suitable solution at the ith iteration, θbest is the solution
8 with the minimum value of D, θworst is the solution with
>
< 0; v < vci ; v > vco
the maximum value of D, θ′j, k, i is the modified of θj, k,
P¼ mv þ c; vci < v < vr (1)
>
: i, and r1, j, i and r2, j, i are two randomly generated vari-
Pmax ; vr < v < vco ables within the interval [0, 1].
4 TECHNICAL NOTE

TABLE 4 RMSEs of the two methods

Standard
Model Min Max Mean Deviation

Linearized 17.615 181.657 48.911 31.494


Logistic 17.829 221.998 41.289 26.686

TABLE 5 MAEs of the two methods

Standard
Model Min Max Mean Deviation

Linearized 12.399 153.948 37.394 24.121


Logistic 13.598 171.298 30.952 19.359

1 N
MAEt ¼ ∑ jPet ðvi Þ − Pat ðiÞj: (7)
N i¼1

RMSEs and MAEs of the two models are calculated


and shown in Tables 4 and 5.
As shown in Tables 4 and 5, it is obviously noticed that
the maximal RMSE and MAE of the linearized segmented
4.3 | Modeling results model are smaller compared with that of the logistic model
whereas the minimal RMSE and MAE show the same ten-
Based on Equations 1 and 3, parameter vectors are
dency. However, the mean and standard deviation values
obtained by using the least square method and Jaya
of RMSEs and MAEs of the linearized segmented model
algorithm, respectively. The statistic result of daily
are higher than those of the logistic model. It is reasonable
parameters of the linearized segmented model is illus-
to conclude that the logistic model describes the data
trated in Table 2, and the statistics of daily parameters
more accurately than the linearized segmented model.
of the four‐parameter logistic model is shown in
Table 3. To assess the effectiveness of two models, the
root mean square error (RMSE) and the mean absolute 5 | ANOMALY DETECTION
error (MAE) are considered as the metrics:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 5.1 | Multivariate approach
1 N
RMSEt ¼ ∑ ðPet ðvi Þ−Pat ðiÞÞ2 (6) In order to monitor the WTPCs, multivariate approach,
N i¼1
including Hotelling's T2 and generalized variance control
charts are implemented.9,20 Based on the linear and logistic
models, model parameters are obtained. Given Xt = (mt, ct)T
TABLE 2 Results of the linearized segmented model
and Yt = (at, mt, nt, τt)T, the corresponding expected value
Standard U and V, the covariance matrix Σ1 for vector Xt, and
Parameter Min Max Mean Deviation the covariance matrix Σ2 for vector Yt are denoted as
m 64.971 278.341 147.005 46.722 " #
−1097.250 −157.253 −475.400 T σ02 σ 01 2
c 195.360 U ¼ ðμ0 ; μ1 Þ Σ1 ¼ (8)
σ 10 2 σ12
TABLE 3 Results of the four‐parameter logistic model

Standard V ¼ ðv0 ; v1 ; v2 ; v3 ÞT (9)


Parameter Min Max Mean Deviation 2 3
σ 2 σ 01 2 σ 02 2 σ 03 2
6 0 7
a 134.096 1769.047 964.135 549.152 6 7
6 σ 10 2 σ 1 2 σ 12 2 σ 13 2 7
m −993.766 924.100 −19.420 205.532 Σ2 ¼ 6
6
7
7
6 σ 20 2 σ 21 2 σ 2 2 7
n 13.758 5000.000 759.305 1459.522 4 σ 23 2 5
τ 0.404 3.202 1.175 0.437 σ 30 2 σ 31 2 σ 32 2 σ3 2
TECHNICAL NOTE 5

FIGURE 2 Phase 1 of linearized parameter vectors monitored with the T2 chart and generalized variance chart [Colour figure can be
viewed at wileyonlinelibrary.com]

FIGURE 3 Phase 2 of linearized parameter vectors monitored with the T2 chart and generalized variance chart [Colour figure can be
viewed at wileyonlinelibrary.com]

where μ0 and μ1 are the mean values of mt and ct; v0, v1, v2, In the linearized segmented model, the statistic of T2
and v3 are the mean values of at, mt, nt, and τt; σ02, σ12, control chart is computed as
σ22, and σ32 are the corresponding variances, while σ012,
σ022, …, σ322 are the covariances. T t 2 ¼ ðX t −U ÞT Σ1 −1 ðX t − U Þ: (10)
6 TECHNICAL NOTE

In the logistic model, the statistic of T2 control chart is where b1 and b2 are obtained as follows:
computed as p
1
b1 ¼ p ∏ ðt − i Þ
T t 2 ¼ ðY t −V ÞT Σ1 −1 ðY t − V Þ: ðt−1Þ i¼1
" #
p p p
(11) 1
b2 ¼ ∏ ðt − iÞ ∏ ðt − j þ 2Þ − ∏ ðn − jÞ ;
The upper control limits (UCL) of Equations 10 and 11 ðt−1Þ2p i¼1 j¼1 j¼1

2ðt − 1Þ 4ðt − 1Þ (13)


are F 2;t−2;α and F 2;t−4;α , respectively.21
t−2 t−4 where p denotes the parameter quantity.
Besides T2 control chart, generalized variance chart is Parameter vectors of two types of models are moni-
applied as well. The center line (CL) and lower control tored by the T2 chart and generalized variance control
limit (LCL) as well as the UCL of GV control chart are chart. Considering the unknown faults of WTs, the
defined as monitoring process is divided into two stages (Stage 1
and Stage 2). At stage 1, control boundaries are derived
CL ¼ b1 jSj
to identify the anomalies initially. At Stage 2, the decision
jSj 1
 limits are re‐calculated by excluding the abnormal points
LCL ¼ b1 − 3b1 2 (12)
b1 detected at Stage 2. Next, the abnormal samples are
rechecked according to the fault log to examine the effec-
jSj 1

tiveness of the proposed approach.
UCL ¼ b1 þ 3b2 2
b1

TABLE 6 The associated date of abnormal samples via the LIN- 5.2 | Monitoring results
EAR model
Because the fault log is only available during the time
Sample Date between 1 April and 14 June, parameter vectors of sample
13 2014/4/15 1 to sample 73 are monitored using the multivariate
51 2014/5/23 control charts.
As shown in Figure 2, samples 51, 55, and 60 of the linear
55 2014/5/27
model are considered to be anomalies. After removing these
60 2014/6/1
data points, the control limits are re‐calculated. It is observed

FIGURE 4 Phase 1 of logistic parameter vectors monitoring with the T2 chart and generalized variance chart [Colour figure can be viewed
at wileyonlinelibrary.com]
TECHNICAL NOTE 7

FIGURE 5 Phase 2 of logistic parameter vectors monitoring with the T2 chart and generalized variance chart [Colour figure can be viewed
at wileyonlinelibrary.com]

that sample 13 is identified as abnormal in Figure 3. In TABLE 7 The associated date of abnormal samples via the
summary, the detected four anomalous samples (13, 51, logistic model
55, 60), and their corresponding date are listed in Table 6.
Sample Date Sample Date
As for parameter vectors obtained from the logistic
model, the Phase 1 results are presented in Figure 4, and 2 2014/4/2 20 2014/4/22
it is notable that more outliers are identified compared with 5 2014/4/3 23 2014/4/25
those of the linear model. As a result, samples 5, 6, 11, 15, 6 2014/4/6 36 2014/5/8
17, 20, and 42 are considered as anomalies. The detection 11 2014/4/13 38 2014/5/10
results of Phase 2 are shown in Figure 5. Samples 2, 23,
15 2014/4/17 42 2014/5/14
36, and 38 are considered as abnormal ones. All associated
dates of these anomalous samples are provided in Table 7. 17 2014/4/19
When comparing Tables 6 and 7, it is observed that con-
trol charts using the logistic model detect more anomalies TABLE 8 Manual examination results of the selected turbine
than those using the linearized model whilst the sequence Time Period Power Curve Analysis Results
numbers of detected samples are different. Moreover, with
1 April‐14 April Impaired
the visual inspection of power curves of selected WTs
(presented in Table 8), the control charts using the logistic 15 April‐19 April Scattered but slightly improved
model demonstrate better performance as the detected 20 April‐24 April Impaired
anomalies generally match the fault log. During the first 25 April‐30 April Scattered but slightly improved
malfunction period, four out of 11 outliers are identified 1 May‐9 May Significantly degraded
using the proposed approach. Although the performance
10 May‐11 May Scattered
of the WT is slightly improved during the next period, two
more anomalies are detected. The propose approach 12 May‐30 May Improved but slightly scattered
successfully monitors two abnormal WTPCs for the time 1 June Scattered
period from 20 April to 30 April. Overall, only one outlier 2 June‐14 June Improved
is undetected during the period of the WT degradation.
The control charts using the linear model fail to find a by using the linear model, the proposed approach is able
number of anomalies, especially during the remarkable to identify most of the abnormal WTPCs. However, the
degradation period. Compared with the detection results SCADA data are collected with an interval of 10 minutes
8 TECHNICAL NOTE

and some unusual patterns are missing. In the future study, 5. Yang W, Court R, Jiang J. Wind turbine condition monitoring by the
higher frequency data need to be considered. approach of SCADA data analysis. Renew Energy. 2013;53:365‐376.
6. Lydia M, Selvakumar AI, Kumar SS, Kumar GEP. Advanced
algorithms for wind turbine power curve modeling. IEEE Trans-
6 | CONCLUSION actions on Sustainable Energy. 2013;4(3):827‐835.
7. Shokrzadeh S, Jozani MJ, Bibeau E. Wind turbine power curve
In this paper, a novel WTPC monitoring approach‐based modeling using advanced parametric and nonparametric methods.
Jaya algorithm was presented. In the proposed approach, IEEE Transactions on Sustainable Energy. 2014;5(4):1262‐1269.
Jaya algorithm was employed to identify parameters of 8. García Márquez FP, Tobias AM, Pinar Pérez JM, Papaelias M.
the four‐parameter logistic model of WTPCs, and the Condition monitoring of wind turbines: techniques and
derived parameters were monitored using the multivariate methods. Renew Energy. 2012;46:169‐178.
control charts. The effectiveness and efficiency of the pro- 9. Long H, Wang L, Zhang Z, Song Z, Xu J. Data‐driven wind tur-
posed monitoring approach were validated with SCADA bine power generation performance monitoring. IEEE
data of WTs from a commercial wind farm in China. Transactions on Industrial Electronics. 2015;62(10):6627‐6635.
The numerical study confirmed that Jaya algorithm 10. Little M, Pope K. Performance modelling for wind turbines
could be used to identify parameters of WTPC models operating in harsh conditions. International Journal of Energy
without using any prior knowledge and algorithm‐specific Research. 2017;41(3):417‐428.
parameters. Meanwhile, more accurate anomaly detection 11. Li S, Haskew TA, Hong Y‐K. Investigation of maximum wind
results were yielded using the four‐parameter logistic power extraction using adaptive virtual lookup‐table approach.
model compared with the linear model employed in the International Journal of Energy Research. 2011;35(11):964‐978.

recent study. Therefore, it is practical to apply the pro- 12. Lei J, Gong Q. Optimal allocation of a hybrid energy storage sys-
posed monitoring approach in real industrial applications. tem considering its dynamic operation characteristics for wind
power applications in active distribution networks. International
Journal of Energy Research. 2018;42(13):4184‐4196.
A C K N O WL E D G E M E N T 13. Ma J, Man KL, Guan S‐U, Ting TO, Wong PWH. Parameter esti-
This work was supported in part by the Guangdong Pro- mation of photovoltaic model via parallel particle swarm
optimization algorithm. International Journal of Energy
vincial Key Laboratory of New and Renewable Research
Research. 2016;40(3):343‐352.
and Development under Grant Y807s61001, in part by
14. Wang Y, Li L. Li‐ion battery dynamics model parameter estima-
the University of Science and Technology Beijing—
tion using datasheets and particle swarm optimization.
National Taipei University of Technology Joint Research International Journal of Energy Research. 2016;40(8):1050‐1061.
Program under Grant TW2018008, in part by the Funda-
15. Rao R. Jaya: a simple and new optimization algorithm for solving
mental Research Funds for the Central Universities constrained and unconstrained optimization problems. Interna-
under Grant 06500078, and in part by the National Natu- tional Journal of Industrial Engineering Computations. 2016;7:19‐34.
ral Science Foundation of China under Grant 71473155. 16. Villanueva D, Feijóo A. Comparison of logistic functions for modeling
wind turbine power curves. Electr Pow Syst Res. 2018;155:281‐288.
ORCID 17. Wang L, Zhang Z, Huang C, Tsui KL. A GPU‐accelerated paral-
lel Jaya algorithm for efficiently estimating Li‐ion battery model
Long Wang https://orcid.org/0000-0001-6695-6054 parameters. Appl Soft Comput. 2018;65:12‐20.
Chao Huang https://orcid.org/0000-0002-8012-3697
18. Huang C, Wang L, Lai L. Data‐driven short‐term solar irradi-
ance forecasting based on information of neighboring sites.
R EF E RE N C E S IEEE Transactions on Industrial Electronics. 2018; in press;1.
19. Huang C, Wang L, Yeung RS, et al. A prediction model‐guided
1. Qi W, Liang Y, Z‐JM S. Joint planning of energy storage and Jaya algorithm for the PV system maximum power point track-
transmission for wind energy generation. Oper Res. 2015;63(6): ing. IEEE Transactions on Sustainable Energy. 2018;9(1):45‐55.
1280‐1293.
20. Kang L, Albin SL. On‐line monitoring when the process yields a
2. Sanchez H, Escobet T, Puig V, et al. Fault diagnosis of an advanced linear profile. Journal of Quality Technology. 2000;32(4):418‐426.
wind turbine benchmark using interval‐based ARRs and observers. 21. Holmes DS, Mergen AE. Improving the performance of the T2
IEEE Transactions on Industrial Electronics. 2015;62:3783‐3793. control chart. Quality Engineering. 1993;5(4):619‐625.
3. Wang L, Zhang Z, Xu J, Liu R. Wind turbine blade breakage
monitoring with deep autoencoders. IEEE Transactions on How to cite this article: Jin R, Wang L, Huang
Smart Grid. 2018;9(4):2824‐2833.
C, Jiang S. Wind turbine generation performance
4. Wang L, Zhang Z, Long H, Xu J, Liu R. Wind turbine gearbox monitoring with Jaya algorithm. Int J Energy Res.
failure identification with deep neural networks. IEEE Transac- 2019;1–8. https://doi.org/10.1002/er.4382
tions on Industrial Informatics. 2017;13(3):1360‐1368.

Das könnte Ihnen auch gefallen