Monitoring Wind Farms With Performance Curves

Transcription

1 192 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 4, NO. 1, JANUARY 2013 Monitoring Wind Farms With Performance Curves Andrew Kusiak, Member, IEEE, and Anoop Verma, Student Member, IEEE Abstract Three different operational curves the power curve, rotor curve, and blade pitch curve are presented for monitoring a wind farm s performance. A five-year historical data set has been assembled for constructing the reference curves of wind power, rotor speed, and blade pitch angle, with wind speed as an input variable. A multivariate outlier detection approach based on -means clustering and Mahalanobis distance is applied to this data to produce a data set for modeling turbines. Kurtosis and skewness of bivariate data are used as metrics to assess the performance of the wind turbines. Performance monitoring of wind turbines is accomplished with the Hotelling control chart. Index Terms Control chart, -means clustering, Mahalanobis distance, performance monitoring, turbine performance curves. I. INTRODUCTION R ECENT years have witnessed a rapid expansion of wind energy. The growing number of wind turbine installations presents the asset operators with turbine operations and maintenance issues [1]. Variable wind conditions contribute to varying loads, which in turn lead to wind turbine faults, e.g., spalled bearings and fractured gears. Failures of wind turbines require repairs and negatively impact their performance [2] [4]. Monitoring the performance of wind turbines is the key to reduction of operations and maintenance costs. Wind turbine performance is affected by: 1) external factors such as wind turbulence, storms, and icing; and 2) internal factors such as components temperature and lubrication. While the external factors cannot be controlled, events associated with the internal factors can be predicted. An efficient way to measure the impact of internal factors is through turbine operations. The operational characteristics of turbines depend on parameters such as rotor power, torque, and pitch angle. Continuous monitoring of these parameters can be useful in assessing wind turbine performance. State-of-the-art turbine monitoring applications are discussed in [3] [5]. Condition monitoring approaches, namely vibration analysis [6], oil analysis [7], and strain measurement [8], are widely used in the literature. Data-mining-based approaches have gained their popularity in wind turbine research. Related applications of data mining include: 1) power curve monitoring [9], 2) monitoring and prediction of wind turbine states [10], Manuscript received January 21, 2012; revised June 01, 2012; accepted August 05, Date of publication September 19, 2012; date of current version December 12, The research reported in this paper was supported by funding from the Iowa Energy Center under Grant The authors are with the Intelligent Systems Laboratory, The University of Iowa, Iowa City, IA USA ( andrew-kusiak@uiowa.edu; anoop-verma@uiowa.edu). Color versions of one or more of the figures in this paper are available online at Digital Object Identifier /TSTE TABLE I DATA DESCRIPTION [11], 3) modeling turbine abnormal behavior [12], [13], and 4) prognosis of turbine faults [14]. In this paper, turbine performance is assessed with three performance curves power curve, rotor curve, and blade pitch curve. To perform such an assessment, operational wind turbine data is needed. Supervisory control and data acquisition (SCADA) systems record wind turbine parameters at different time intervals. SCADA data may be effectively used to tune a wind farm and provide early warnings of possible failures. In the research reported in this paper, historical wind turbine data is used to extract reference curves. The paper is organized as follows. Section II presents the data set used in the research, along with a description of turbine performance curves. In Section III, the solution methodology is presented, which consist of: 1) reference curve construction, 2) outlier detection, and 3) moment calculation. In Section IV, a statistical control chart is employed for performance monitoring of wind turbines. Section V concludes the paper with topics for future research. II. DATA FOR TURBINE PERFORMANCE CURVES The data used in this paper has been collected from a wind farm of over 100 wind turbines. Three wind turbine performance curves power curve (power versus wind speed), rotor curve (rotor speed versus wind speed), and blade pitch curve (blade pitch angle versus wind speed) are constructed for wind turbine performance. The data set analyzed in this paper is divided into three parts. First, a four-year historical data (August 2005 August 2008) from 22 turbines is available to extract the reference curves for the month of August. The data is averaged over 10-min intervals (10-min data). The reference curves are validated with data from the following year (August 2009). To conduct performance monitoring, one month of data from 22 wind turbines was collected in August Table I presents the data used in the research. Descriptions of the three turbine performance curves are provided in Section II-A. A. Wind Turbine Performance Curves Apowercurve indicates power generated by a wind turbine at various wind speeds. Malfunctions of a wind turbine will impact its power-generation capability. A typical wind power curve resembles a sigmoid function; however, due to various malfunctions (e.g., sensors and components), the power curve acquires ashape of its own /$ IEEE

2 KUSIAK AND VERMA: MONITORING WIND FARMS WITH PERFORMANCE CURVES 193 Fig. 1. Proposed solution methodology. A rotor curve represents a mapping between rotor speed and wind speed. Failures of turbine components impact its shape. A typical rotor curve is a monotonically increasing function of the wind speed. A blade pitch curve shows the relationship between the turbine pitch angle and wind speed. The turbine s control system adjusts the blade angle for maximum power capture. A malfunction of the control system and high wind speed causes a turbine to stall, i.e., blade pitch angle becomes 90. During the normal operations of a wind turbine, the pitch angle is set to, e.g., 0, 66, and 83. In general, during the startup of a wind turbine, the blade pitch is set to a high value. A negative value of the pitch angle reflects the presence of a strong wind. In the cut-in to cut-out region of the wind speed, the blade pitch settings are adjusted by the control system for the maximum power output. At the rated wind speed, the blade pitch angle is continuously adjusted to maintain the power required. In Section III, the proposed solution methodology is discussed. Fig. 2. Average monthly wind speed distribution near the wind farm location (source: Iowa Energy Center). III. SOLUTION METHODOLOGY The proposed solution methodology includes four phases (see Fig. 1). The historical wind farm data from several wind turbines is scanned initially to select wind turbine data (Phase 1). Due to the stochastic nature of the wind and inherent variability in the individual turbines, the noisy raw data is processed using a multivariate outlier detection approach (Phase 2). The resulting reference curves are used as a benchmark to evaluate the performance of individual turbines (Phase 3). Skewness and kurtosis of performance curves are calculated for each wind turbine and compared against the corresponding reference curves. In Phase 4, a quality control chart is used for performance monitoring of wind turbines. All solution phases are discussed in the subsequent sections. A. Construction of Reference Curves Wind turbine performance depends on wind speed. The data used in this research is obtained from a large wind farm located in Blairsburg, Iowa. The monthly average distribution of wind speed at this location is investigated. Fig. 2 provides the monthly distribution of wind speed. The average wind speed varies across different months. Constructing reference curves using the yearly performance data may not be ideal as many details may be missed. Based on the distribution of the wind speed, the reference curves for individual months are constructed. Based on the completeness of data across different years, performance curves for the month of August are extracted from the turbine data collected on 22 turbines over a four-year period (i.e., August 2005 August 2008). Based on the analysis, Fig. 3. Performance curves for the month of August (August 2005 August 2008). (a) Power curve. (b) Rotor curve. (c) Blade pitch curve. data from 22 wind turbines is averaged to obtain three reference curves. Fig. 3(a) (c) provides the reference power curve, rotor curve, and blade pitch curve of the historical monthly data for the month of August. In Section III-B, a bivariate outlier selection approach is discussed. B. Bivariate Outlier Detection The reference curves constructed from the historical data contain outliers which need to be removed for clear depiction of

3 194 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 4, NO. 1, JANUARY 2013 normal turbine behavior. These outliers are largely due to the sensor errors and fluctuations in the turbine performance. In this paper, a multivariate outlier detection approach based on Mahalanobis distance is used. The Mahalanobis metric expresses the distance of an instance to the centroid in the multidimensional space [15], and it is calculated based on the correlation-covariance matrix. Therefore, Mahalanobis distance indicates whether an instance is an outlier with respect to the independent variable values In (1), is the Mahalanobis distance between instance and,and is the inverse of covariance matrix. Due to distinct shape of performance curves, calculating Mahalanobis distance for an overall curve can be misleading as the centroid (usually for wind speeds between 4.5 and 7 m/s) will consider the extreme data points (points close to cut-in wind speed, or/and near rated wind speed) as outliers, which in fact they are not. Thus, in order to improve outlier detection, the performance curve data is grouped into smaller clusters. The -means clustering algorithm determines the number of clusters for each curve by minimizing the cost function in the following [16]: where is the clustering cost, is the number of clusters, is the number of data points in cluster, represents the data points, and represents cluster. The proposed procedure for identifying outliers in the bivariate performance curves is presented next. Procedure: Extracting Smooth Performance Curves Parameters:, optimal number of clusters (kopt), maximum number of data subsets (Max_fold), Mahalanobis distance threshold Begin For each Set initial number of clusters DividethedatasetintoMax_fold For Randomly select 90% subset for training and 10% for testing. Initialize centroids Repeat, until the centroid does not change Evaluate the training error using the cost function in (2) (1) (2) End End for Output: For (Optimal number of clusters) Evaluate the Mahalanobis distance data pair from the cluster mean Sortthedatapairs Retain data pairs End for based on the distance with Output: Mahalanobis distance between all data points for clusters Repeat Do until Output: Smooth performance curves (1) for each In the above procedure, the Mahalanobis distances between the data points and the cluster centers (centroids) are computed. The -means clustering algorithm applied to the monthly reference curves provides 14, 11, and 9 clusters for the power curve, the rotor curve, and the blade pitch curve, respectively. Fig. 4(a) (c) illustrates the clustered reference curves with training errors 0.039, 0.062, and for these same curves, respectively. Fig. 5 depicts the Mahalanobis distance of power curve for individual clusters. The outlier data points can be easily identified in Fig. 5. Similarly, rotor curve and blade pitch curve are analyzed. The aim here is to extract smooth reference curves; therefore, a conservative approach based on the Mahalanobis distance metric is used to remove the outlier data points. The threshold distance is chosen in such a way that the data points corresponding to the high-density clouds in the clusters are selected. Table II presents the threshold distance for each cluster. Using the threshold distance indicated in Table II, 10% 15% of the data points were considered as outliers and thus are discarded. The refined performance curves are illustrated in Fig. 6(a) (c). In Section III-C, the moment of the performance curves is discussed. Scalar performance matrices, namely, skewness and kurtosis of bivariate data, are evaluated. C. Moment Calculation The third- and fourth-order moments, namely kurtosis and skewness, are often used to describe the shape of the data distribution. The multivariate kurtosis and skewness can also be used as a data compression technique providing a single value describing the shape of the distribution. Due to the high frequency data that is used in the proposed research, the performance curves are described using kurtosis and skewness. The refined reference curves obtained in Section III-B are used as a

4 KUSIAK AND VERMA: MONITORING WIND FARMS WITH PERFORMANCE CURVES 195 Fig. 5. Mahalanobis distance (MD) of power-curve-based clusters. (a) Clusters (b) Clusters tivariate data, where a value close to zero indicates elliptical symmetry. The multivariate skewness is defined [17], [18] Fig. 4. Performance curves with clusters. (a) Power curve. (b) Rotor curve. (c) Blade pitch curve. (3) where is the matrix mean, and is the estimated population covariance matrix. Similarly, multivariate kurtosis is a univariate measure of kurtosis for multivariate data. For column matrix, a value of kurtosis coefficient close to indicates approximate multinormality. Multivariate kurtosis is mathematically described in the following [18], [19]: (4) benchmark for performance monitoring of the wind farm. Multivariate skewness is a univariate measure of skewness for mul- In general, under-performing wind turbines will deviate from the reference curves, resulting in different values of kurtosis and

5 196 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 4, NO. 1, JANUARY 2013 TABLE II MAHALANOBIS DISTANCE THRESHOLD FOR PERFORMANCE CURVE CLUSTERS TABLE III MULTIVARIATE KURTOSIS AND SKEWNESS OF REFERENCE CURVES skewness that can be tracked in a 2-D graph. Depending on the requirements, performance monitoring can take place on a daily, weekly, or monthly basis. Table III compares the kurtosis and skewness of yearly and monthly reference curves with respect to the test data (August 2009). The kurtosis and skewness of reference curves constructed based on the yearly data (January 2008 December 2008) is obtained in a similar way. The values presented in Table III indicate that the skewness and kurtosis of monthly reference curves are much closer than of the yearly reference curves. Thus, monthly reference curves are used for monitoring the wind farm. D. MonitoringaWindFarm In this section, data from 22 wind turbines over a period of a month (August 2011) are analyzed using kurtosis and skewness for three performance curves. The analysis is based on the 10-min average data. Turbines located farthest from the reference points (see Table III) are considered to be abnormal. Euclidean distance is used to evaluate the distance of individual wind turbines from the reference points. Figs. 7 9 provide a 2-D scatter plot of the performance curves, where each point (diamond) represents an individual wind turbine. In addition, kurtosis and skewness of the reference curves are included in Figs Depending on the distribution of data points across the performance curves, the kurtosis and skewness distribution varies. Due to the distinct shape of the power curves, the kurtosis and skewness values are higher and more spread out than those of the rotor and blade pitch curves. In the 2-D skewness-kurtosis graph, wind turbine performance can be assessed by: 1) relative location of individual turbines with respect to the reference curves, and 2) location of individual turbines with respect to the turbine clusters. In Fig. 6. Refined performance curves. (a) Power curve. (b) Rotor curve. (c) Blade pitch curve. general, turbine showing the same behavior will form a distinct cluster. Any abnormal turbine behavior can be easily visualized in a 2-D scatter graph. The possible reasons for the distinct location of individual turbines in the skewness-kurtosis plot could be: 1) under-performance due to system abnormalities, 2) under-performance due to different wind speeds, and 3) over-performance due to errors in wind speed measurement.

6 KUSIAK AND VERMA: MONITORING WIND FARMS WITH PERFORMANCE CURVES 197 Fig. 7. Status of a wind farm reflected by the power curve. Fig. 9. Status of a wind farm reflected by the blade pitch curve. Fig. 10. Power curve of turbine showing abnormal behavior (turbine 10). Fig. 8. Status of a wind farm reflected by the rotor curve. Using the guidelines mentioned earlier in this section, the power-curve-based skewness-kurtosis graph identifies turbines 10 and 12 as abnormal, whereas turbines 6, 9, 13, and 17 behave differently in the rotor and blade pitch curves. Fig. 10 illustrates the power curve of turbine 10. The abnormal behavior of turbine 10 is clearly visible as the fault logs confirm the faults associated with generator windings. More information about turbine fault logs is provided in [3] and [20]. Performance of wind turbines can be assessed with the 2-D kurtosis-skewness graph; however, time is not depicted in the scatter plot. Therefore, to keep track of time, control charts are utilized. IV. PERFORMANCE MONITORING OF WIND TURBINES In this section, performance monitoring of wind turbines is performed using a quality control chart. The overall monitoring of the wind farm can be done on a weekly or monthly basis; however, for performance monitoring of a wind farm over time, quality control charts are required. In this paper, two output metrics are used skewness and kurtosis. Monitoring the matrices independently can be misleading. Therefore, bivariate process monitoring using Hotelling s control chart is employed. In the literature, Hotelling s chart has been widely used to simultaneously monitor two or more output variables [21]. Equations (5) (7) define the statistic as follows: In (5), is the individual observation, is the mean, is the covariance matrix inverse, and is the index of input variables. Since the subgroup size is 1, the covariance matrix is evaluated by pooling all observations [22], [23] For two output variables, the covariance matrix will take the form The lower control limit (LCL) is always 0, whereas the upper control limit (UCL) is calculated from (7) In (8), is the number of output variables, is obtained from distribution. The value of is set to The kurtosis (5) (6) (7) (8)

7 198 IEEE TRANSACTIONS ON SUSTAINABLE ENERGY, VOL. 4, NO. 1, JANUARY 2013 Fig. 11. The Box-Cox transformation of turbine 1 data. Fig. 12. Control limits for training data points (iteration 1). Fig. 13. Control limits for training data points (iteration 3). and the skewness can be monitored simultaneously; however, the Hotelling test requires the data to be normal. Therefore, the initial data is normalized using the Box-Cox approach. A value varies from 5.0to5.0.Fig.11providesacomparison of initial and transformed skewness data of turbine 1 obtained using the Box-Cox approach. For equal to 0.552, the transformed data resembles a normal distribution. This process is repeated for all turbines. The transformed bivariate data of turbines is divided into two parts, e.g., training and testing. Using the information presented in the power-curve-based kurtosisskewness data (see Fig. 7), turbines 7, 10, 11, 12, and 15 are used for testing, whereas the control limits are obtained by using data from the remaining turbines. Fig. 12 provides the UCL of , resulting in 12 data points out of control. The out-ofcontrol data points are removed, and the training process is iterated until all data-points meet the control limits. After three training process iterations, all data points were found in control, with the resulting UCL of (Fig. 13). Fig. 14(a) (e) illustrates the test data corresponding to turbines 7, 10, 11, 12, and 15. Based on the obtained UCL values, turbines 7, and 11 were found to be affected by an abnormal day Fig. 14. Hotelling s chart for the test data (a) turbine 7, (b) turbine 10, (c) turbine 11, (d) turbine 12, and (e) turbine 15. in a month, whereas, two abnormal days were found for turbine 10. The remaining turbines, e.g., turbines 12 and 15 were found to operate normally. In general, various faults (cable twisting left, faulty pitch controller) contributed to the abnormal days of turbines 7, 10, and 11. However, the main reason for abnormal days is the power curtailment. Due to the limited number of observations, no significant pattern in the value was observed. V. CONCLUSION A systematic approach for monitoring the performance of a wind farm was presented. Three performance curves the power curve, rotor curve, and blade pitch curves were used. The Mahalanobis distance was computedtoidentifyoutliersin the performance curves. The bivariate performance curve data was grouped into several clusters for better identification of outliers. Using the skewness and kurtosis of bivariate data, the

8 KUSIAK AND VERMA: MONITORING WIND FARMS WITH PERFORMANCE CURVES 199 initial high-frequency data was compressed to a single value. Hotelling s control chart was used for performance monitoring of the data points in time. The transformed kurtosis-skewness graphs were determined to be better suited for turbine monitoring than the high-frequency performance curves. The Hotelling control chart was applied to the daily average data. Future research will focus on hourly data. The impact of the output variables on Hotelling s control chart deserves further study. REFERENCES [1] A.Zaher,S.D.McArthur,andD.G.Infield, Online wind turbine fault detection through automated SCADA data analysis, Wind Energy, vol. 12, no. 6, pp , [2] Z.Hameed,Y.S.Hong,Y.M.Cho,S.H.Ahn,andC.K.Song, Condition monitoring and fault detection of wind turbines and related algorithms: A review, Renew. Sustain. Energy Rev., vol. 13, no. 1, pp. 1 39, [3] R. Hyers, J. McGowan, K. Sullivan, J. Manwell, and B. Syrett, Condition monitoring and prognosis of utility scale wind turbines, Energy Mater., vol. 1, no. 3, pp , [4] Y.Amirat,M.E.H.Benbouzid,E.Al-Ahmar,B.Bensaker,andR. Wamkeue, A brief status on condition monitoring and fault diagnosis in wind energy conversion systems, Renew. Sustain. Energy Rev., vol. 13, no. 9, pp , [5] P. Tavner, G. W. Bussel, and F. Spinato, Machine and converter reliabilities in wind turbines, in Proc. 3rd IET Int. Conf. Power Electronics Machines and Drives, Ireland, 2006, pp [6] P. Caselitz and J. Giebhardt, Rotor condition monitoring for improved operational safety of offshore wind energy converters, J. Solar Energy Eng., vol. 127, no. 2, pp , [7] E. Becker and P. Posta, Keeping the blades turning: Condition monitoring of wind turbine gears, Refocus, vol. 7, no. 2, pp , [8] T.W.Verbruggen,WindTurbineOperationandMaintenanceBasedon Condition Monitoring,WT-, Final Report Energy Research Center, The Netherlands, ECN-C , 2003, pp [9] A. Kusiak, H.-Y. Zheng, and Z. Song, On-line monitoring of power curves, Renew. Energy, vol. 34, no. 6, pp , [10] A. Kusiak and A. Verma, Prediction of status patterns of wind turbines: A data-mining approach, ASME J. Solar Eng., vol. 133, no. 1, pp , [11] A. Kusiak and A. Verma, A data-mining approach to monitoring wind turbines, IEEE Trans. Sustain. Energy, vol. 3, no. 1, pp , Jan [12] A. Zaher and S. McArthur, A multi-agent fault detection system for wind turbine defect recognition and diagnosis, in Proc. IEEE, Lausanne, Switzerland, Jul. 2007, pp , POWERTECH. [13] M. A. Sanz-Bobi, M. C. Garcia, and J. D. Pico, SIMAP: Intelligent system for predictive maintenance: Application to the health condition monitoring of a wind turbine gearbox, Comput. Industry, vol. 57, no. 6, pp , [14] A. Kusiak and W. Y. Li, The prediction and diagnosis of wind turbine faults, Renew. Energy, vol. 36, no. 1, pp , [15] P. Mahalanobis, On the generalized distance in statistics, Proc. National Institute of Sciences of India, vol. 2, no. 1, pp , [16] M. Inaba, N. Katoh, and H. H. Imai, Applications of weighted voronoi diagrams and randomization to variance-based K-clustering, in Proc. 10th ACM Symp. Computational Geometry, 1994, pp [17] K. V. Mardia, Measures of multivariate skewness and kurtosis with applications, Biometrika, vol. 57, no. 3, pp , [18] K. V. Mardia, Applications of some measures of multivariate skewness and kurtosis in testing normality and robustness studies, Sankhya, ser. B, vol. 36, no. 2, pp , [19] K. V. Mardia, Tests of Univariate and Multivariate Normality, in Handbook of Statistics, P. Krishnaiah, Ed. Amsterdam, The Netherlands: North Holland, 1980, vol. 1, pp [20] A. Kusiak and A. Verma, Enhanced turbine performance monitoring, Wind Syst. Mag., vol. 3, no. 24, pp , [21] R. A. Johnson and D. W. Wichern, Applied Multivariate Statistical Analysis. Upper Saddle River, NJ: Prentice-Hall, [22] J. D. Williams, W. H. Woodall, J. B. Birch, and J. H. Sullivan, Distribution of Hotelling s T2 statistic based on successive difference estimator, J. Quality Technol., vol. 38, no. 3, pp , [23] J. H. Sullivan and W. H. Woodall, A comparison of multivariate quality control charts for individual observations, J. Quality Technol., vol. 28, no. 4, pp , Andrew Kusiak (M 90) received the B.S. and M.S. degrees in engineering from the Warsaw University of Technology, Warsaw, Poland, in 1972 and 1974, respectively, and the Ph.D. degree in operations research from the Polish Academy of Sciences, Warsaw, in He is Professor and Chair of the Department of Mechanical and Industrial Engineering, University of Iowa, Iowa City. His current research interests include applications of computational intelligence in automation, wind energy, manufacturing, product development, and healthcare. Prof. Kusiak is the Institute of Industrial Engineers Fellow and the Editor-in ChiefoftheJournal of Intelligent Manufacturing. Anoop Verma (S 10) received the B.S. degree in manufacturing engineering from the National Institute of Foundry and Forge Technology, Ranchi, Jharkhand, India, and the M.S. degree in mechanical engineering from the University of Cincinnati, Cincinnati, OH, in 2007 and 2009, respectively. He is currently working toward the Ph.D. degree in the Industrial Engineering Program, The University of Iowa, Iowa City. He has researched on areas such as machine loading, manufacturing scheduling, corporate memory, rapid prototyping, and information retrieval, and published papers in various journals. He reviewed papers for different journals, including the Computer Aided Design, International Journal of Computer Integrated Manufacturing, International Journal of System Science, Information Science, and Journal of Intelligent Manufacturing. His current doctoral research focuses on applications of data mining and computational intelligence in wind industry and wastewater treatment facilities.