Data-Driven Wind Turbine Power Generation ...

This is the author's version of an article that has been published in this journal. Changes were made to this version by the publisher prior to publication. The final version of record is available at http://dx.doi.org/10.1109/TIE.2015.2447508

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

1

Data-Driven Wind Turbine Power Generation Performance Monitoring Huan Long, Long Wang, Zijun Zhang, Member, IEEE, Zhe Song, and Jia Xu  Abstract—This research investigates the wind turbine power generation performance monitoring based on SCADA data. The proposed approach identifies turbines with weakened power generation performance through assessing the wind power curve profiles. Profiles which statistically summarize the curvatures and shapes of a wind power curve over consecutive time intervals are constructed by fitting power curve models into SCADA datasets with a least square method. To monitor variations of wind power curve profiles over time, the multivariate and residual approaches are introduced and applied. Two blind industrial studies are conducted to validate the effectiveness of the proposed monitoring approach and the study results demonstrate a high accuracy of detecting abnormal power curve profiles of wind turbines as well as their associated time intervals. Index Terms—Performance monitoring; wind energy; power curve; multivariate approach; residual analysis

I. INTRODUCTION

T

ECHNOLOGICAL maturity and foreseeable improvement of cost competitiveness made the wind energy a widely accepted energy solution of sustainable living and limited supply of fossil fuels. The wind power industry has rapidly expanded during the past decades and the hotspot of wind energy technology development has gradually shifted from improving the wind power generation efficiency [1 – 3] to serving the wind farm operations and maintenance (O&M) at present [4, 5]. Advancing wind farm O&M technologies including condition monitoring is of great interests to wind farm operators because renewable energy subsidies are not sustainable. In the literature, studies of wind farm condition monitoring majorly could be categorized as frequency domain analysis and Manuscript received September 28, 2014; revised March 28, 2015 and May 17, 2015; accepted May 21, 2015. Copyright (c) 2015 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to [email protected]. The work described in this paper was supported in part by an Early Career Scheme grant from the Research Grants Council of the Hong Kong Special Administrative Region (Project No. CityU 138313) and in part by the National Natural Science Foundation of China (Grant No. 71001050). Huan Long, Long Wang and Zijun Zhang are with the Department of Systems Engineering and Engineering Management, City University of Hong Kong, Hong Kong, China (e-mail: [email protected]). Zhe Song is with the School of Business, Nanjing University, Nanjing, China (e-mail: [email protected]). Jia Xu is with the Centre of the Wind Farm Data Analysis and Performance Optimization at China Longyuan Power Group Co. Ltd., Beijing, China (e-mail: [email protected])

time domain analysis. Frequency domain analysis studies examined the power spectrum of frequencies of vibration and acoustic emission signals and aimed to reasoning root causes of wind turbine sub-system abnormal conditions [6 – 9]. Wavelet transformation and its extended methods have been widely applied. Tang et al. [10] applied Morlet wavelet transformation to filter vibration signal noise and a Wigner-Ville distribution based time-frequency analysis to identify faults. Tsai et al. [11] presented a continuous wavelet transform-based approach to identify impaired conditions of wind turbine blades. Watson et al. [12] presented a wavelet-based monitoring approach to analyze high frequent power output data. The frequency domain analysis techniques are powerful to diagnose faults of individual wind turbine assemblies; however, it requires extra instruments to collect high frequent data which can induce additional cost and system complexity. Time domain analyses focused on detecting wind turbine abnormal statuses through tracking trends of condition parameter values [13-15]. The power generation performance of wind turbines can be characterized by its power curves constructed from real-time SCADA (Supervisory Control and Data Acquisition) data. Thus, in the time domain analyses, previous wind turbine condition monitoring studies focused on analyzing the quality of produced power curves. Kusiak et al. [16] introduced non-parametric methods for modeling the wind power curve from industrial data and analyzed model fitting residuals to detect anomalies. The approach presented in [16] was then extended to realize the online monitoring [17]. Kusiak and Li [18] investigated fault diagnosis through analyzing patterns of power curve fitting residuals. The philosophy of studies [16-18] is to develop statistical boundaries for detecting outliers based on a derived reference power curve. The similar approach has been applied to monitor vibration of the wind turbine gearbox [19]. Besides residual analyses, Qiu et al. [20] and Feng et al. [21] presented physics-based data analysis methods for detecting anomalies of wind turbine assemblies and performing diagnosis with SCADA data. Compared with the frequency domain analysis, the time domain analyses are directly conducted based on the collected SCADA data. The condition monitoring and fault diagnosis have both been studies with the time domain analysis. Based on the long-term observation of wind power curves, people recognized that the power curve curvature and shape of a healthy wind turbine is variable rather than fixed. Gill et al. [21] and Wang et al. [22] applied Copula-based approaches to conduct initial analyses of the relationships of multiple wind

Copyright (c) 2015 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].



2

turbine parameters and detect their variations. The reported monitoring approaches [16-18] based on fixed reference power curves did not count the contribution of wind power curve variations and thus could produce false alarms. The wind power curve can shift due to factors rather than faults, such as, the change of air density. Thus, it is challenging to determine the wind turbine conditions with analyzing individual data points. In this research, a monitoring approach is proposed to monitor the wind power generation performance through analyzing the variation of wind power curves rather than individual data points. The SCADA data of a wind turbine is partitioned into sub-datasets based on consecutive equal time intervals and each sub-dataset provides a power curve of the wind turbine over the time interval. The linear and Weibull CDF based power curve models are applied to generate power curve profiles by fitting into the sub-datasets. Power curve profiles are composed of power curve model parameters which depict power curve characteristics. A multivariate approach is applied to monitor multiple parameters simultaneously and a residual approach is utilized to analyze residuals of fitting power curves. The effectiveness of the presented approach is demonstrated through two blind industrial studies.

the profile for depicting the characteristic of the wind power curve contained in 𝒮 is x if and only if x satisfies (1).

II. WIND POWER CURVE PROFILE In reality, the power curve of a wind turbine is dynamic due to variations of factors including weather, air density, system controls, etc. Let t denote M time windows with equal lengths, t = 1, 2, …, M, the power curves of a wind turbine in different time windows deviate as shown in Fig. 1. It is observable that the power curve shifts over time when the wind turbine is healthy. In addition, the power curve 2 includes many outliers. A traditional monitoring approach for analyzing individual points will generate a set of alarms for the power curve 2 which is not necessary. To generate more reliable monitoring results, a data-driven profile monitoring approach is proposed. 1600

Power outpur (kW)

1400 1200

N

x  arg min{ ( Pn  f (vn , x)) 2 }

(1)

n 1

Let t index 𝒮, t = 1, 2, …, M, a set of power curve profiles, xt, can be obtained through (1) for monitoring. Numerous power curve modeling approaches have been reported in [21-23]. In this study, we select two types of profiles, the linear profile and Weibull cumulative distribution function (WCDF) profile, to demonstrate the accuracy and effectiveness of the proposed monitoring framework. A. Linear profile The fundamental approach for approximating a wind power curve is a stepwise linear model shown in (2). vci  v, v  vco 0,  P  av  b, vci  v  vr (2) P , vr  v  vco  max where a is the slope, b is the intercept, Pmax is the maximal power capacity, vci is the cut-in wind speed, vr is the rate wind speed, and vco is the cut-out wind speed. Based on (2), it is observable that characteristics of wind power curves are depicted by both a and b. Through fitting model (2) to 𝒮t by (1), the linear power curve profiles are xt = T

(at, bt) where at and bt yield the least square error for fitting (2) to 𝒮t. B. Weibull cumulative distribution function profile Since the shape of Weibull CDF function and wind power curve is highly similar and the variable in Weibull distribution is naturally positive, wind power curves can be described by the Weibull CDF based stepwise model in (3). vci  v , v  vco 0,    xc (3) P   Pmax 1  e   , vci  v  vr  vr  v  vco   Pmax ,



k



where k and c are shape and scale parameters of the Weibull distribution. Other notations have been presented in Section T II.A. The same as Section II.A, the WCDF profiles, xt = (kt, ct) , are generated through fitting model (3) to 𝒮t.

1000 800 600 400 200

0 0

2

4

6 8 Wind speed (m/s)

Power curve 1

10

12

14

Power curve 2

Fig. 1. The wind power curves of a turbine in two different time windows.

The target monitored in the proposed monitoring framework includes the wind power output, P, and the wind power curve profile described in Definition 1. Definition 1. Given a data set, 𝒮 = {(v1, P1), (v2, P2), …, (vN, PN)}, where v is the wind speed, P is the power output, and N is the number of data points, and a model, f(x), for fitting 𝒮 with a least square method, where x is a vector of parameters in f(∙),

III. MONITORING APPROACHES The wind power curve monitoring can be decomposed to three scenarios according to vci, vr, and vco. When v < vci and v > vco, wind turbines do not produce power and the power curve monitoring is not needed in this scenario. When vr ≤ v ≤ vco, the wind power is rated at Pmax. To monitor turbine conditions, we can simply apply the line P = Pmax to examine the deviation of observed power outputs from Pmax. When vci ≤ v < vr, the shape and curvature of the power curve of a healthy wind turbine is variable over time (see Fig. 1). To monitor the quality of power curves in such scenario, profiles are developed as described in Section II to conclude characteristics of power curves based on




3

sub-datasets. Impaired power curves can demonstrate abnormal curvature and shape caused by different reasons as shown in Fig. 2 and 3. Fig. 2 demonstrates the abnormal curvature of power curves might be caused by sub-system malfunctions, severe weather conditions, power curtailment commands, etc. Fig. 3 shows the abnormal shape of power curves possibly due to the system degradation and inappropriate control settings. To identify the power curve with an abnormal curvature and shape, a multivariate approach for monitoring power curve profiles and a residual approach for monitoring power curve fitting errors are introduced in next sections.

 12 are the corresponding variance, and the  012 is the covariance. 2 Based on (4), the sample statistic in T control chart is computed as (5).

Tt 2   xt  μ  Σ1  xt  μ  T

The T control chart only has upper control limit (UCL) and it is  2,2  when μ and Σ are known. However, in reality, the μ and Σ are typically unavailable. Thus, the X and S, the unbiased estimates of μ and Σ, defined in (6) need to be computed.

Power output (kW)

1550 1350

1150

   ,   

  v2 1 v2 ) MSE   MSE(  N Svv  Svv  S 2 MSE  MSE  v   S Svv  vv  

950 750 550 350 150 0

5

10 Wind speed (m/s)

15

20

(6)

      

where MSE and Svv are calculated as (7) and (8). M

Abnormal power curve

Normal power curve

MSE 

Fig. 2. Power curves with abnormal and normal curvatures.

N

  N  2  ( P  f (v , x)) 1

t 1

n 1

n

2

n

(7)

M N

Svv   (vn  v )2

1800 1600

Power output (kW)

T

M  M   x0,t  x1,t X   t 1 , t 1 M  M  

1750

-50

(5)

2

(8)

n 1

1400

2

The sample statistic for the T control chart is then described in (9). M T (9) T 2   xt  X S1  xt  X  M 1 2 The UCL of T control chart is then equal to 2F2,(N – 2)M, α [20].

1200

1000 800 600

400 200 0

2

0

5

10 Wind speed (m/s)

Abnormal power curve

15

Besides the T control chart, the generalized variance chart with the UCL, central line (CL), and lower control limit (LCL) expressed in (10) is also conducted to examine the shift of Σ.

20

Normal power curve

UCL 

Fig. 3. Power curves with abnormal and normal shapes.

A. The Multivariate Approach As presented in (2) and (3), the linear and Weibull CDF based power curve profiles are vectors of two parameters. A straightforward solution of monitoring profiles is to construct control charts to monitor two parameters separately. However, such solution presumes the independency between two parameters and omits the correlation. It is obvious that two parameters in (2) and (3) are dependent. To monitor two 2 parameters simultaneously, the Hotelling’s T control chart [20, 21] is applied. T Given profiles, xt = (x0,t, x1,t) , t = 1, 2, …, M, where x0,t is at in (2) and kt in (3) while x1,t is bt in (2) and ct in (3), the expected value μ and the variance-covariance matrix Σ of xt are defined as (4). 2   2  01  μ  ( 0 , 1 )T , Σ   02 2    1   01

(4)

where the μ0 and μ1 are the mean of x0,t and x1,t, ∀t, the  02 and

|S|  B1  3B20.5  , B1

(10)

CL  B1 | S |, LCL 

|S|  B1  3B20.5  B1

where B1 and B2 are computed as (11) and (12) given the number of parameters q. B1 

B2 

1 ( N  1) q

q

 ( N  i)

(11)

i 1

q q  q 1 ( N  i ) ( N  j  2)  (N     ( N  1)2 q i 1 j 1  j 1

 j) 

(12)

B. The Residual Approach The multivariate approach is effective to detect power curves with irregular curvature but might not be sensitive to the variation of power curve shapes. For example, as shown in Fig. 3, it is obvious that a similar pair of x will be generated if we fit either model (2) or model (3) to the data. The residuals of model fitting can be analyzed to detect power curves with




4

impaired shapes. The I-MR chart [21] composed of two charts, the individual chart and moving range (MR) chart, is applied to perform monitoring. The integration of individual and MR charts offers two dimensions of residual monitoring information: 1) the trend of model fitting residuals; 2) the variation between model fitting residuals. Let et denote the mean absolute percentage error (MAPE) of fitting model f(∙) to datasets, 𝒮t, t = 1, 2, …, M, expressed in

10-s SCADA data of wind turbines collected from a 100 MW Class wind farm in U.S. The second study investigates the 10-min wind turbine SCADA data collected from a wind farm owned by China Long Yuan Power Group Corporation Limited. In two blind studies, the proposed approach was firstly applied without the prior knowledge of considered wind turbines to generate the monitoring results. Next, the results of the first blind study are validated with the fault logs while the results of the second study are examined by the domain experts in China Long Yuan Power Group Corporation.

(13), the averages of et, e , for t = 1, 2, …, M, can be obtained through (14). N

et  1 N   Pn  f (vn ) Pn

(13)

n 1

M

e   et M

(14)

t 1

To construct I-MR chart for et, the MRt needs to be computed based on (15). (15) MR t  et  et 1 Let MR be the average of MRt, the UCL, CL, and LCL of individual chart are presented in (16). UCL  e  3

MR MR , CL  e , LCL  e  3 d2 d2

(16)

where d2 is 1.128 because we consider the moving range of 2 consecutive observations [21]. To construct the MR chart, its UCL, CL, and LCL are set up as (17). (17) UCL  D4 MR, CL  MR, LCL  D3 MR where D4 and D3 are constants depending on n. Since n = 2 in this study, the D4 = 3.267 and D3 = 0 according to [21]. C. The Online Monitoring Procedure The online monitoring procedure of the proposed approach shown in Fig. 4 is composed of four steps:

ate ari ltiv ach Mu ppro a

Sub-dataset 1 Sub-dataset 2 Partition

⁞ Sub-dataset N

SCADA Data Collection

Sub-datasets Generation

Transform

Re app sidua roa l Profiles ch Development

A. Study One In study one, the vci, vr, and vco of the considered wind turbine are 3.5 m/s, 11.5 m/s, and 20 m/s separately. The monitoring results are compared with the fault logs to show the detection accuracy. 1. Data description In this study, two months 10-s data collected from Jan 1st, 2011 to Feb 28th, 2011 from a wind turbine in a U.S. wind farm were applied to the power curve profile construction and the condition monitoring. The collected data contained values of wind turbine performance parameters, wind conditions, as well as the fault logs. SCADA data are partitioned into 177 sub-files based on the fixed time interval and each sub-file included 3600 data points which recorded 10-h power generation performance of the wind turbine. 2. Power curve profile computation The least square method is applied to fitting model (2) and (3) based on data of 177 sub-files to generate two types of power curve profiles. The ranges of c and k for fitting model (3) are set to 1 ≤ c ≤ 100 and 1 ≤ k ≤ 5. The range of c is determined artificially. The minimum and maximum of k are set to 1 and 5 because k > 5 will lead to P = Pmax when v < vr and k < 1 will lead to a concave curve for Weibull CDF rather than an s-shape curve. The generated power curve profiles are summarized in Table I and II. TABLE I STATISTICAL SUMMARY OF LINEAR PROFILES

Fig. 4. The online monitoring procedure of the proposed model.

Step 1. Partition the SCADA dataset to sub-datasets according to fixed time intervals. Step 2. Generate power curve profiles and power curve fitting errors based on sub-datasets. Step 3. Apply multivariate approach and residual approach to develop control charts for monitoring Step 4. Continuously collect SCADA data, produce new power curve profiles, and apply developed control charts IV. BLIND INDUSTRIAL STUDIES This section presents two blind industrial studies based on the proposed approach. The first study is conducted based on

TABLE II STATISTICAL SUMMARY OF WEIBULL CDF BASED PROFILES

In Table I, it is observable that the values of a and b in linear profiles have significant variations. However, the median and mean indicate that a and b are more likely to be positive and negative among all linear profiles. Table II summarizes the values of c and k in WCDF profiles statistically. It is observable that the values of c and k have the possibility of reaching the




5

minimal and maximal boundary. The median and mean of c and k reflect that they should normally be around 10 and 3.6 separately. The fitting residuals of models (2) and (3) are statistically summarized in Table III and IV. The RMSE is the square root

detected in Phase 1 and 2 analyses are provided in Table V. Through monitoring WCDF profiles, the T2 and GV chart detect that wind turbine power curves are abnormal when t = 2, 3, 16, 21, 47, 76, 77, 94, 98, 99, 154, 155, 158, 159, and 174.

N

of

T squared chart of slope, intercept

1  N  2   ( Pn  f (vn , x))2 and the MAPE is computed

50

n 1

based on (13). TABLE III FITTING RESULTS OF LINEAR PROFILES

T squared

40

30

20 UCL=17.17 10 Median=2.11 0

TABLE IV FITTING RESULTS OF WEIBULL CDF BASED PROFILES

1

19

37

55

73

91

109

127

145

163

Profiles

Fig. 5. The Phase 1 monitoring of linear profiles with T2 chart. Generalized variance chart of slope, intercept 7

UCL=4.275

4 3 2

|S|=1.309

0

LCL=0 1

19

37

55

73

91

109

127

145

163

Profiles

Fig. 6. The Phase 1 monitoring of linear profiles with GV chart. Tsquared chart of slope, intercept 60 50 40

T squared

3. Monitoring with the multivariate approach The monitoring process is split to two phases, Phase 1 and 2, because the prior information of the wind turbine faults is unknown. In Phase 1, the control limits are developed to filter outliers. In Phase 2, the control limits are recomputed after removing outliers to perform monitoring. Next, abnormal profiles detected by the multivariate and residual approaches are analyzed by checking with the fault logs. The multivariate approach is firstly applied to monitor the linear profiles generated in Section IV. In the Phase 1 analysis, 2 the T and generalized variance (GV) charts are developed in Figs. 5 and 6 to filter the outliers. It is observable that the profiles 2, 94, 98, 99, and 105 are 2 considered as outliers by T chart while the GV chart provides the similar result. Thus, the information of profiles 2, 94, 98, 2 99, and 105 is eliminated and the control limits of T and GV chart are re-calculated. The monitoring results in Phase 2 are demonstrated in Fig. 7 and 8. In Phase 2 analysis, the wind turbine condition in the profile 100 is considered as abnormal because the sample statistic 2 exceeds the limit of T chart. In conclusion, the wind turbine power curves in total 6 linear profiles (2, 94, 98, 99, 100, and 105) are considered as abnormal. The similar monitoring procedure is applied to examine the WCDF profiles. The index of abnormal power curve profiles

5

1

30 20 UCL=17.15 10 Median=2.11

0 1

18

35

52

69

86

103

120

137

154

171

Profiles

2

Fig. 7. The Phase 2 monitoring of linear profiles with T chart. Generalized variance chart of slope, intercept 5 UCL=4.412 4

Generalized variance

Based on the RMSE of Table III and IV, we observe that the minimal and maximal RMSE provided by linear models are less than those provided by Weibull CDF based model in model fittings. However, the median, mean, and standard deviation of RMSE offered by Weibull CDF based model are significantly lower than those of linear models which indicate that general performance of fitting wind power curves based on Weibull CDF based models is slightly better than that based on linear models. It is reasonable because the curvature of Weibull CDF is more close to that of the reference wind power curve.

Generalized variance

6

3

2 |S|=1.351 1

0

LCL=0 1

18

35

52

69

86

103

120

137

154

171

Profiles

Fig. 8. The Phase 2 monitoring of linear profiles with GV chart.




6

TABLE V MONITORING RESULTS OF WCDF PROFILES WITH MULTIVARIATE APPROACH

TABLE VI MONITORING RESULTS OF WCDF PROFILES WITH RESIDUAL APPROACH

4. Monitoring with the residual approach The abnormal profiles detected by the multivariate approach were excluded before applying the residual approach. The condition monitoring with the residual approach also includes two phases. The Phase 1 is applied to filter outliers and the Phase 2 aims to detect anomalies of power curves. The Phase 1 individual and MR charts based on linear profiles are demonstrated in Fig. 9. The top chart in Fig. 9 is the individual chart while the bottom one is the MR chart. Four profiles at t = 34, 82, 87, and, 137 with MAPE higher than 1000% are considered as outliers and filtered. After filtering outliers in Phase 1, the control limits of individual and MR charts are re-computed to conduct the Phase 2 analysis based on linear profiles presented in Figs. 10. Table VI presents the profiles with irregular fitting errors with the residual approach in phase 1 and 2 based on WCDF profiles. All the detected abnormal residuals of profiles include profiles at t = 27, 34, 38, 82, 83, 106, 108, 109, 111,137, 148, 157, 172, and 177.

5. The analysis of detected faults Through the eye observation, we discover that power curves based on 27 sub-datasets display abnormal curvatures. The multivariate approach successfully identified 17 of the 27 profiles based on linear and WCDF profiles. The patterns of 27 irregular profiles can be briefly categorized to three types as shown in Figs. 11 – 13 based on the root causes.

Fig. 11. Power curve of sub-dataset 2.

I-MR Chart of MAPE 1

4800

Individual Value

1

3600 1

2400

1

1200 UCL=615 _ X=117 LCL=-381

0 1

17

33

49

65

81

97

113

129

145

P rofiles 11

4800

Moving Range

11

3600 11

2400

11

1200 UCL=612 __ MR=187 LCL=0

0 1

17

33

49

65

81

97

113

129

145

P rofiles

Fig. 9. The Phase 1 monitoring of linear profiles with the residual approach.


I-MR Chart of MAPE 1 11

Individual Value

400

1

1

300

1 1

200

1

1

1

UCL=156.1 100

_ X=43.7

0 LCL=-68.7 1

17

33

49

65

81

97

113

129

145

P rofiles

1

400

1

Moving Range

1 1

1

11

300 1

200

11

1

1

1

1

UCL=138.1 100

__ MR=42.3 LCL=0

0 1

17

33

49

65

81

97

113

129

145


P rofiles

Fig. 10. The Phase 2 monitoring of linear profiles with the residual approach.

The anomaly of power curve in Fig. 11 is majorly caused by the power curtailment command although pitch faults, low




7

gearbox oil pressure, and battery problems were reported in the log. In Fig. 12, the wind turbine does not generate power but draws power from the grid because the wind speed is lower than the cut-in speed most of the time. The power curve in Fig. 13 is abnormal because during the period the wind turbine suffered a series faults including the yaw problem, diverter malfunction, incorrect blade angle, generator brushes worn, cable twisting, and periodical re-startup. The unidentified profiles through the multivariate approach generally displayed a pattern similar to the Fig. 14. In sub-dataset 27, the wind speed fluctuated around the cut-in wind speed and it is difficult to control the wind turbine to produce a stable power curve.

the considered wind turbines are 3m/s, 10m/s, and 21m/s separately. The collected SCADA data is partitioned into sub-datasets with 2-day intervals for generating the power curve profiles and each sub-dataset includes 288 data points. The proposed monitoring approach has been applied to monitor the power generation performance of all 32 wind turbines. Since the fault logs are not available and industrial partner only manually examined power generation performance of two wind turbines, the proposed approach is validated through comparing generated monitoring results and the industrial analysis report of two selected wind turbines. The same monitoring procedure in study one is applied here. The linear and Weibull CDF based models are applied to generate power curve profiles based on sub-datasets covering 10-min SCADA data of 2 days. The multivariate and residual approaches are employed to monitor the linear and Weibull CDF based power curve profiles as well as their fitting errors. The monitoring results of selected wind turbine 1 and 2 are presented in Table VII and VIII separately. The analysis results of domain experts are provided in Table IX. TABLE VII MONITORING RESULTS OF SELECTED WIND TURBINE 1


The residual approach detects 5 of the left 10 visually detected abnormal profiles because those profiles also lead to significant residuals of fitting models. The more significant contribution of applying the residual approach is to detect 10 more profiles with the pattern of power curve shown in Fig. 15. Although the power curve in Fig. 15 displays a normal curvature, data points reflect that wind turbine was temporarily shut down many times when v > vci. The log recorded that the wind turbine experienced faults including emergency stop, brake malfunction, as well as generator brush worn and received a repair during the operating period.

TABLE VIII MONITORING RESULTS OF SELECTED WIND TURBINE 2

TABLE IX INDUSTRIAL ANALYSIS RESULTS OF TWO TURBINES

Fig.15. Power curve with the normal curvature but impaired shape.

B. Study Two In this study, 10-min SCADA data of a wind farm includes 32 1.6 MW Class wind turbines in China collected from April 1st, 2014 to July 30th, 2014 is investigated. The vci, vr, and vco of

The domain experts only examined two wind turbines for a period from April 1st – June 1st. Thus, the monitoring results generated with the proposed approach over the same period are selected and reported in Table VII and VIII. In Table VII and VIII, symbols, a, b, c, and d, are applied to label abnormal power curves detected by the multivariate approach based on linear profiles, the residual approach based on linear profiles, the multivariate approach based on WCDF profiles, and the




8

residual approach based on WCDF profiles. Table IX reports the brief summary of manual examinations of two selected wind turbine which provided the similar power generation performance. Through comparing Tables VII and VIII with Table IX, we could observe that the monitoring results generated by the proposed approach could basically match the summary of the industrial examination report. Since such monitoring study was conducted based on 10-min data, better results could be produced if a higher frequent data, such as 10-s data, are available.

Detection,” IEEE Trans. Ind. Electron., Vol. 62, No. 3, pp. 1693-1702, 2015. W. Wang and P. McFadden, “Application of wavelets to gearbox vibration signals for fault detection,” Journal of Sound and Vibration, Vol. 192, pp. 927-939, 1996. X. Gong and W. Qiao, “Bearing Fault Diagnosis for Direct-Drive Wind Turbines via Current-Demodulated Signals,” IEEE Trans. Ind. Electron., Vol. 60, No. 8, pp. 3419-3428, 2013. K. Iyer, X. Lu, Y. Usama, V. Ramakrishnan, and N. Kar, “A Twofold Daubechies-Wavelet-Based Model for Fault Detection and Voltage Regulation in SEIGs for Distributed Wind Power Generation,” IEEE Trans. Ind. Electron., Vol. 60, No. 4, pp. 1638-1651, 2013. B. Tang, W. Liu, and T. Song, “Wind turbine fault diagnosis based on Morlet wavelet transformation and Wigner-Ville distribution,” Renewable Energy, Vol. 35, No. 12, pp. 2862-2866, 2010. C. Tsai, C. Hsieh, and S. Huang, “Enhancement of damage-detection of wind turbine blades via CWT-based approaches,” IEEE Trans. on Energy Convers., Vol. 21, No. 3, pp. 776-781, 2006. S. Watson, B. Xiang, W. Yang, P. Tavner, and C. Crabtree, “Condition Monitoring of the Power Output of Wind Turbine Generators Using Wavelet,” IEEE Trans. Energy Convers., Vol. 25, No. 3, pp. 715-721, 2010. H. Sanchez, T. Escobet, V. Puig, and P. Odgaard, “Fault Diagnosis of an Advanced Wind Turbine Benchmark Using Interval-Based ARRs and Observers,” IEEE Trans. Ind. Electron., Vol. 62, No. 6, pp. 3783-3793, 2015. J. Blesa, P. Jimenez, D. Rotondo, F. Nejjari, and V. Puig, “An Interval NLPV Parity Equations Approach for Fault Detection and Isolation of a Wind Farm,” IEEE Trans. Ind. Electron., Vo. 62, No. 6, pp. 3794-3805, 2015. N. Freire, J. Estima, and A. Marques Cardoso, “Open-Circuit Fault Diagnosis in PMSG Drives for Wind Turbine Applications,” IEEE Trans. Ind. Electron., Vol. 60, No. 9, pp. 3957-3967, 2013 A. Kusiak, H. Zheng, and Z. Song, “Models for monitoring wind farm power,” Renewable Energy, Vol. 34, No. 3, pp. 583-590, 2009. A. Kusiak, H. Zheng, and Z. Song, “On-line monitoring of power curves,” Renewable Energy, Vol. 34, No. 6, pp. 1487-1493, 2009. A. Kusiak and W. Li, “The prediction and diagnosis of wind turbine faults,” Renewable Energy, Vol. 36, No. 1, pp. 16-23, 2011. Z. Zhang, A. Verma, and A. Kusiak, “Fault Analysis and Condition Monitoring of the Wind Turbine Gearbox,” IEEE Trans. Energy Convers., Vol. 27, No. 2, pp. 526-535, 2012. Y. Feng, Y. Qiu, C. Crabtree, H. Long, and P. Tavner, “Monitoring wind turbine gearboxes,” Wind Energy, Vol. 16, No. 5, pp. 728-740, 2013. Y. Qiu, W. Zhang, M. Cao, Y. Feng, and D. Infield, “An Electro-Thermal Analysis of a Variable-Speed Doubly-Fed Induction Generator in a Wind Turbine,” Energies, Vol. 8, No. 5, pp. 3386-3402, 2015. S. Gill, B. Stephen, and S. Galloway, “Wind Turbine Condition Assessment Through Power Curve Copula Modeling,” IEEE Trans. Sustain. Energy, Vol. 3, No. 1, pp. 94-101, 2012. Y. Wang, D. Infield, B. Stephen, and S. Galloway, “Copula-based model for wind turbine power curve outlier rejection,” Wind Energy, Vol. 17, No. , pp. 1677-1688, 2014. Z. Zhang, Q. Zhou, and A. Kusiak, “Optimization of Wind Power and Its Variability With a Computational Intelligence Approach,” IEEE Trans. Sustain. Energy, Vol. 5, No. 1, pp. 228-236, 2014. S. Li, D. Wunsch, E. O’Hair, and M. Giesselmann, “Comparative Analysis of Regression and Artificial Neural Network Models for Wind Turbine Power Curve Estimation,” ASME Journal of Solar Energy Engineering, Vol. 123, No. 4, pp. 327-332, 2001. T. Ustuntas and A. Sahin, “Wind turbine power curve estimation based on cluster center fuzzy logic modeling,” Journal of Wind Engineering and Industrial Aerodynamics, Vol. 96, No. 5, pp. 611-620, 2008. L. Kang and S. Albin, “On-line Monitoring When the Process Yields a Linear Profile,” Journal of Quality Technology, Vol. 32, No. 4, pp. 418-426, 2000. D. Montgomery, Introduction to Statistical Quality Control. John Wiley & Sonc, Inc., 2008.

[7]

[8]

[9]

[10]

V. CONCLUSION A data-driven approach for monitoring the variation of a wind turbine power curve was presented to identify impaired power generation performance. To depict the curvature and shape of a wind power curve, the power curve profiles were developed. Two power curve models, the linear and Weibull CDF based model, were applied to generate the linear and WCDF power curve profiles based on a sequence of SCADA sub-datasets. In the monitoring, a multivariate approach was applied to analyze power curve profiles while a residual approach was utilized to analyze errors produced by fitting power curve models. Two blind industrial studies were conducted to prove the effectiveness and accuracy of the proposed monitoring approach. The results demonstrated that good monitoring accuracy was obtained and few false alarms were generated. It was because that the monitoring analysis of the proposed approach was conducted based on sub-datasets rather than individual points. The proposed approach could be implemented for online monitoring. It could detect faults accurately but not as timely as the point-based condition monitoring methods. It is simply because a certain amount of time was required for constructing new power curve profiles. The future research will investigate a mechanism to partition datasets with the completion of the power curve information to enhance the monitoring power of the proposed approach. In addition, the relationship between wind power and multiple parameters will be considered in the monitoring. REFERENCES [1]

[2]

[3]

[4]

[5]

[6]

R. Li and D. Xu, “Parallel Operation of Full Power Converters in Permanent-Magnet Direct-Drive Wind Power Generation System,” IEEE Trans. Ind. Electron., Vol. 60, No. 4, pp. 1619-1629, 2013. J. Chen, J. Chen, and C. Gong, “New Overall Power Control Strategy for Variable-Speed Fixed-Pitch Wind Turbines Within the Whole Wind Velocity Range,” IEEE Trans. Ind. Electron., Vol. 60, No. 7, pp. 2652-2660, 2013. A. Kusiak and Z. Zhang, “Control of wind turbine power and vibration with a data-driven approach,” Renewable Energy, Vol. 43, pp. 73-82, 2012. A. Kusiak, Z. Zhang, and G. Xu, “Minimization of Wind Farm Operational Cost Based on Data-Driven Models,” IEEE Trans. Sustain. Energy, Vol. 4, No. 3, pp. 756-764, 2013. S. Bernal-Perez, S. Ano-Villalba, R. Blasco-Gimenez, and J. Rodriguez-D’Derlee, “Efficiency and Fault Ride-Through Performance of a Diode-Rectifier- and VSC-Inverter-Based HVDC Link for Offshore Wind Farms,” IEEE Trans. Ind. Electron., Vol. 60, No. 6, pp. 2401-2409, 2013 X. Gong and W. Qiao, “Current-Based Mechanical Fault Detection for Direct-Drive Wind Turbines via Synchronous Sampling and Impulse

[11]

[12]

[13]

[14]

[15]

[16] [17] [18] [19]

[20] [21]

[22]

[23]

[24]

[25]

[26]

[27]

[28]




9

Huan Long received the B.E. degrees in Automation from Huazhong University of Science and Technology, Wuhan, China, in 2013. She is currently working towards the Ph.D. in the Department of Systems Engineering and Engineering Management at the City University of Hong Kong, Hong Kong, China. Her research interests include the wind farm layout planning & analysis, wind turbine monitoring, as well as wind farm performance robust optimization.

Long Wang received his M. S. degree with Distinction in Computer Science from University College London, London, England, UK, in 2014, M. Eng. degree in Hydraulic Engineering from China Agricultural University, Beijing, China, in 2013 and B. Eng. degree in Irrigation and Drainage Engineering from China Agricultural University, Beijing, China, in 2011. He is currently pursuing his Ph. D. degree in the Department of Systems Engineering and Engineering Management at the City University of Hong Kong, Hong Kong, China. Zijun Zhang (M’12) received his Ph.D. and M.S. degrees in Industrial Engineering from the University of Iowa, Iowa City, IA, USA, in 2012 and 2009, respectively, and B.Eng. degree in Systems Engineering and Engineering Management from the Chinese University of Hong Kong, Hong Kong, China, in 2008. Currently, he is an Assistant Professor in the Department of Systems Engineering and Engineering Management at the City University of Hong Kong, Hong Kong, China. His research focuses on data mining and computational intelligence with applications in modeling, monitoring, optimization and operations of systems in wind energy, HVAC and wastewater processing domains.

Zhe Song received his Ph.D. in Industrial Engineering from the University of Iowa in 2008. After graduation, he continued his research as a postdoctoral researcher at the University of Iowa. He joined the School of Business at Nanjing University as an Associate Professor in 2009. His research interests includes Operations research, Data mining, Control theory, Computational intelligence, and their applications in business, energy and manufacturing systems modeling and optimization including Mass customization, Power plant performance optimization, Wind power management, Optimal decision making in Heating Ventilating and Air Conditioning (HVAC) systems.

Jia Xu received his M.S. degree in Instrument Science and Technology at Xi’an Jiaotong University, China, in 2010. Now, he is the head of the Centre of the Wind Farm Data Analysis and Performance Optimization at China Longyuan Power Group Co. Ltd., Beijing, China.
