A Compound Approach for Monitoring the Variation in Wind Turbine Power Performance with SCADA Data

Abstract

The performance of wind turbines directly determines the profitability of wind farms. However, the complex environmental conditions and influences of various uncertain factors make it difficult to accurately assess and monitor the actual power generation performance of wind turbines. A data-driven approach is proposed to intelligently monitor the power generation performance evolution of wind turbines based on operational data. Considering the inherent nonlinearity and structural complexity of wind turbine systems, a data-derived characteristic construction and dimensionality reduction method based on KPCA is adopted as a prerequisite. Additionally, an AdaBoost-enhanced regressor is applied to wind power prediction with adequate inputs, and day-oriented deviation indicators are further constructed for quantifying performance fluctuations. The final validation phase includes two application cases: In the first case, the results show that the proposed method is sensitive enough to capture the early characteristics of blade damage faults. In the second case, an uncertainty error within ±0.5% demonstrates that the proposed method has high-level accuracy in the quantitative assessment of the power performance and good practical effectiveness in real engineering applications.

1. Introduction

Compared with traditional energy sources, renewable energy, such as wind power, is becoming increasingly popular owing to its cleaner and more efficient nature [1,2]. As an efficient energy source with no carbon emissions and wide distribution around the world, wind energy has gradually developed into one of the most important renewable energy sources in the form of clean energy. In the process of transitioning to low-carbon electricity generation, expanding the scale of wind farms is inevitable. Wind turbines are the core equipment, capturing wind energy and converting it to electricity. The accurate assessment of wind turbines’ power generation performance is crucial for the operation of wind farms. The assessment and real-time monitoring of the power generation performance of wind turbines are further complicated by the multifarious and dynamic external environment and self-imposed defects. It is a matter of formality for operators and manufacturers in the wind power industrial chain to constantly think about how to identify defects and deeply excavate the potential power generation capacity of wind turbines in a timely manner. Supervisory control and data acquisition (SCADA) systems have been widely employed for the operational monitoring of modern wind turbines [3]. As frontier approaches to Industry 4.0, big data, big-data mining, and artificial intelligence techniques bring breakthrough development opportunities to the wind power industry [4].

The continuous monitoring of wind turbine conditions to detect potential failures at an early stage draws attention worldwide. Early failure detection enables prompt responses prior to catastrophic accidents and reduces economic losses and maintenance costs [5]. Benefiting from the boom in cloud computing and data-mining technologies, data analysis methods have been extensively investigated in fault detection and the condition-based monitoring of wind turbine components, such as blades, gearboxes, and bearings, drawing increasing attention to the reliability improvements in wind turbines [6,7]. S. Sun et al. [8] primarily discussed the design of convolutional and recurrent neural networks for extracting spatial and temporal features from SCADA systems. A coarse learner and multiple fine learners were established to enhance the reliability of fault diagnosis results through collective decision making. Y. Miao et al. [9] proposed an improved maximum correlated kurtosis inverse convolution adjustment method for wind turbine gear fault diagnosis based on encoder signals. Convolutional adjustment definitions were introduced to overcome the limitations of traditional deconvolution methods. Finally, real experimental cases validated that encoder signals can serve as alternative tools for wind turbine gear fault diagnosis. L. Wang et al. [10] developed a wind turbine gearbox fault detection method based on a deep, fully connected neural network. S. Sun et al. [11] developed an adaptive method to identify blade damage based on a microphone array and compressive beamforming. A generalized minimax concave penalty function was employed to enhance sparse recovery capacities, and the potential damage locations were extracted in coarse acoustic maps to improve the convergence rates. The experimental results demonstrated that several instances of damage to operating blades can be precisely recognized with high efficiencies, and the deterioration of acoustic maps induced by improper step sizes can be avoided. H. Badihi et al. [12] presented an adaptive proportional integral algorithm with an automatic signal correction system for the detection and diagnosis of pitch controller and pitch actuator faults. S. Cho et al. [13] combined Kalman filtering and an artificial neural network for diagnosing blade pitch system faults and achieved the goal of high diagnostic accuracy in fault instances of wind turbines. A. Dibaj et al. [14] devised a vibration-based fault detection method for offshore wind turbine drivetrains based on the optimal selection of the acceleration measurements. C. Zhang et al. [15] utilized a new method based on a Bayesian and an adapted Kalman-augmented Lagrangian for filtering signals under time-varying conditions for the fault detection of wind turbine blade bearings. A. Wang et al. [16] proposed a novel method and embedded a multivariable query pattern in a self-attention mechanism for quantifying the influences of different features on the anomalies of wind turbines. The experimental results confirmed the effectiveness of the proposed method for the early detection of wind turbine faults and identification of anomaly causes. Z. Wang et al. [17] presented a SCADA data-driven method for the fault diagnosis of wind turbines based on incremental learning and the multivariate state estimation technique (MSET). The experimental results demonstrated that incremental MSET can obtain higher estimation accuracies and lower false-alarm rates in long-term operation and detect potential gearbox faults from hours to weeks in advance.

Meanwhile, the performance assessment of wind turbines mainly relies on power curve modeling, which is a conventionally recognized standard in the field of wind energy [18]. Many scholars have conducted extensive research on power curve modeling, which can be classified into two categories: parametric methods and non-parametric methods [19]. A. Quraan et al. [20] organized a comprehensive review comparing different wind turbine power curve modeling approaches. Several real wind turbines with their experimental data were used in that study. The obtained results showed a good correlation between the estimated energy output and the measured energy output. Y. Wang et al. [21] emphasized the importance of detecting and removing outliers from a dataset before modeling, which improved the accuracy of their power curve assessments. The outliers were directly related to the abnormal operation states of wind turbines, such as power limitations, shutdowns caused by failures and maintenance, and measurement errors. W. Wang et al. [22] proposed an efficient combinational algorithm based on an Isolation Forest (I-Forest) and a mean-shift model for data cleaning in the power curve modeling of wind turbines. Local abnormal data were detected by I-Forest, and the remaining stacked data were eliminated by the mean-shift model. The numerical results of application cases positively confirmed the reliability of the universal framework provided by the proposed algorithm. Q. Yao et al. [23] constructed a model-data hybrid-driven outlier detection method for power curve modeling and designed an adaptive update rule for major parameters in that method to enhance the detection accuracy. E. Gonzalez et al. [24] employed a methodology for wind turbine performance monitoring using high-frequency SCADA data by multivariate-featured and non-parametric methods, and detected the performance degradation through analysis of the power-predicted residuals. P. Guo et al. [25] invoked a Dirichlet Process Gaussian Mixture model for data clustering to identify the contour of the main power band used as a baseline performance region. The Mahalanobis distance was used to determine whether new monitoring data lay outside the contour of the main power band. In the industrial performance degradation case caused by the high temperature of the gearbox oil, the proposed method quickly issued an alarm only 2 hours later than the first degraded operational data appeared. C. Zuo et al. [26] proposed a three-stage multiple data pre-processing algorithm for power curve modeling and introduced a novel evaluation method based on Energy Characteristic Consistency to assess the reliability of these algorithms. Four wind turbines were selected to verify the effectiveness of the nine data processing methods. The results showed that the Kernel Density Estimation with Least Square Method exhibited commendable performance with the smallest cumulative evaluation index values. C. Paik et al. [27] proposed a procedure for outlier elimination based on vector quantization and density-based spatial clustering algorithms in estimating the power curves of wind farms. The proposed methodology was demonstrated to be highly efficient in the presence of noises.

Power curve modeling for wind turbines’ power performance monitoring is intuitive but has limited capability in capturing the precise correspondence between wind speed input and active power output due to its insufficient nonlinear fitting abilities, since the averaging method dilutes the volatility and diversity of operational data. Power curves are unable to provide adequate quantitative information for power performance optimization or degradation monitoring because of the ±5% inherent uncertainty according to IEC61400-12-1 standards [28]. Overall, the research on power performance monitoring technologies is not systematic and in depth enough, and implementable technical solutions are lacking.

In this study, an effective power performance variation tendency monitoring method independent from the power curve is proposed. The operational data after wind turbine grid connection are concentrated. The abnormal operating conditions, such as shutdown data, fault data, and power limit data, are thoroughly cleaned through conditional filtering. After data acquisition and pre-processing, feature construction and data dimensionality reduction are performed based on the existing parameter variables. The kernel principal component analysis (KPCA) method is applied in extracting the top-nine principal components with the Cumulative Percent Variance (CPV) exceeding 96%. Then, a high-accuracy strong learner combined with the AdaBoost enhancement algorithm is employed to construct a power prediction model. The experimental results indicate that the accuracy can be improved by relying on adequate feature construction and regression reinforcement. Finally, the effectiveness of the proposed method is demonstrated through two application cases of wind turbines with performance fluctuations.

2. Data Pre-Processing Methodology

2.1. Data Cleaning and Feature Construction

SCADA systems can collect multi-source heterogeneous data for the operational management of wind turbines, including but not limited to operating state data, environmental data, grid side data, and fault records. These data are transmitted to the intermediate central server via a local area network [29]. Researchers have proposed various SCADA data pre-processing methods, such as data filtering, averaging, and kernel density-based clustering algorithms [30,31]. Following extensive data cleaning, the SCADA data are utilized to perform operational state assessments and the power curve modeling of wind turbines. Before delving into factors, such as sensor accuracy, storage exceptions, communication issues, and system failures, it is crucial to emphasize the significance of data preparation and filtering prior to data mining [32]. This study employs a three-step process to clean the raw data, which includes the following:

Step 1: Eliminate null values, data beyond reasonable ranges, and other invalid data.

Step 2: Remove data from periods of shutdown, power limitation, and fault occurrences according to the operation status and maintenance records.

Step 3: Exclude slipped outliers based on a density-based clustering algorithm DBSCAN.

A total of 26,280 records are collected from a 2.0 MW wind turbine over a six-month period. Following data cleaning, 17,890 valid records remain, as illustrated in Figure 1. It can be observed from the figure that the data in shutdown status, fault, and power limitation status, especially the outlier data, have been accurately identified. The residual data, assumed to be within normal operational status, will be further analyzed to evaluate the power performance of wind turbines in detail.

Gathering the relevant parameters is essential to characterizing the degradation in performance. An initial set of variables, including wind speed, ambient temperature, rotor speed, pitch angle, yaw error, and generator temperature, along with an additional 22 variables, was identified. The raw data were processed to create new derived variables using methods, such as moving average, moving standard deviation, and temporal differencing. This processing employs sliding window widths of 10 minutes, 30 minutes, and 1 hour. Following feature construction, variables that are presumed to affect output power, including both basic and derived variables, are incorporated into the pool of input variables. Subsequently, any invalid data introduced during the feature construction are eliminated repeatedly.

The feature construction process yields a total of 28 basic variables and 28 × 3 derived variables, which serve as input variables, thereby effectively preserving the potential correlations between environmental factors, unit-specific characteristics, and the power performance. This holds substantial importance in enhancing the accuracy of wind turbine power performance assessments.

2.2. Feature Extraction and Dimensionality Reduction

2.2.1. KPCA Methodology

During the data pre-processing stage, feature extraction and data dimensionality reduction play a crucial role in improving the model’s representation capability. Feature extraction involves generating or identifying distinctive features from the set of input variables, while data dimensionality reduction aims to streamline the number of features without compromising accuracy. Addressing the limitation of principal component analysis algorithms in handling nonlinear datasets, K. R. Müller et al. [33] proposed a kernel principal component analysis (KPCA) method based on the original algorithm. A kernel function was employed to map nonlinear data from the original space to a high-dimensional space; then, linear PCA was utilized to extract the underlying nonlinear correlations between variables for achieving data dimensionality reduction [34].

Assume that there is a training data sample $X={\left[\begin{array}{cccc}{x}_1,& x_2,& \cdots ,& x_N\end{array}\right]}^T\in R^{N\times M}$, where M is the number of dimensions and N is the number of data points. Define the nonlinear mapping function: $\mathsf{\Phi }\left(x\right)\in F$, $\mathsf{\Phi }:R^M\to F$, where F is the eigenspace. The covariance matrix of F space is calculated as follows:

$$C^F=\frac{1}{N}{\sum }_{j=1}^N\mathsf{\Phi }\left(X_j\right){\mathsf{\Phi }\left(X_j\right)}^T$$

(1)

In this study, the Gaussian radial basis function $k\left(x,y\right)=e^{\frac{{‖x-y‖}^2}{{2\sigma }^2}}$ is selected as the kernel function for eigenvalue decomposition of Equation (1):

$$\lambda V=C^{F}V=\frac{1}{N}{\sum }_{j=1}^N〈{\mathsf{\Phi }\left(X_j\right)}^T,V〉\mathsf{\Phi }\left(X_j\right)\lambda V$$

(2)

where < > denotes the dot product operator; λ, V are the eigenvalues and eigenvectors of the covariance matrix C^F, $V={\sum }_{j=1}^N\mathsf{\Phi }\left(X_j\right)$. Additionally, the kernel function matrix K$\in R^{N\times M}$, which satisfies the condition $K_{ij}=〈\mathsf{\Phi }\left(X_i\right),\mathsf{\Phi }\left(X_j\right)〉$, is introduced; it is incorporated into the following expression:

$$\frac{1}{N}{\sum }_{i=1}^N{\alpha }_i〈\mathsf{\Phi }\left(X_k\right),{\sum }_{j=1}^N\mathsf{\Phi }\left(X_j\right)〉〈\mathsf{\Phi }\left(X_j\right),\mathsf{\Phi }\left(X_i\right)〉=\frac{1}{N}{\sum }_{i=1}^N{\alpha }_i{\sum }_{i=1}^N{K}_{kj}{K}_{ji}$$

(3)

The data $\mathsf{\Phi }\left(x\right)$ are mapped to the feature vector V_k, and the principal element t is given by:

$$t_k=〈V_k,\mathsf{\Phi }\left(x\right)〉={\sum }_{i=1}^N{\alpha }_i^k〈\mathsf{\Phi }\left(x_i\right),\mathsf{\Phi }\left(x\right)〉k=1,\cdots ,p$$

(4)

where p is the number of nuclear primitives.

The formula to calculate the Cumulative Percent Variance (CPV) based on the cumulative contribution of variance is as follows:

$$CPV\left(p\right)={\sum }_{i=1}^p{\lambda }_i/{\sum }_{i=1}^M{\lambda }_i\times 100\%$$

(5)

where λ_i is the eigenvalue of the covariance matrix, and M is the number of all variables.

2.2.2. Application Example

Similarly, the operational data of the 2.0 MW wind turbine mentioned in Section 2.1 are used as a case study. Data pre-processing is conducted according to the method previously described. The 28 basic variables and 28 × 7 derived variables are initially selected. Then, the KPCA algorithm is applied for data dimensionality reduction, using the mean method to optimize the parameter σ through cross-validation.

In the practical application, the KPCA parameter σ is set to 600. Following the data dimensionality reduction, the original variables are efficiently combined or integrated. The relationship among the first three principal components is depicted in Figure 2. The CPV for the first nine principal elements reaches 96.6%, as shown in Figure 3. This indicates a high level of information retention, allowing for a reasonable explanation of the data characteristics.

3. Power Performance Monitoring Model

3.1. AdaBoost Methodology

Since the 1990s, numerous researchers have been devoted to theoretical and algorithmic advancements in integrated learning. M. Kearns and L. Valiant [35] first introduced the concepts of strong learnability and weak learnability in 1988, and later in 1997, Schapire and Freund [36] demonstrated that these concepts are equivalently efficient for machine learning. The learning task was shifted from constructing a single strong learner with high accuracy to boosting a series of weak learners with moderate accuracy. This led to the development of the adaptive boosting (AdaBoost) algorithm, which employs an iterative adaptation process for incremental improvement.

As depicted in Figure 4, the procedure for power assessment using the AdaBoost strong regressor algorithm is outlined as follows:

Step 1: Initialize sample weights. Select m groups of training data from sample space beforehand. Set the initial weights for each group as follows:

$$\begin{array}{cc}{w}_1\left(i\right)=\frac{1}{m}& i=1,2,\cdots ,m\end{array}$$

(6)

where m is the number of weak regressors. Assuming the maximum number of iterations is N, the initial iteration cycle is set as k = 1.

Step 2: Create weak regressors. The weak regressors are created using the CART decision tree model. During the training of the k’th regressor G_k(x), both the training error ε_{k, i} and the prediction error rate p_k are calculated as follows:

$${\epsilon }_{k,i}=\left|G_k\left(x_i\right)-y_i\right|$$

(7)

$$p_k={\displaystyle \sum _i{\omega }_k\left(i\right)|{\epsilon }_{k,i}>\theta }$$

(8)

where θ is the error threshold.

Step 3: Calculate regressor weights. The weight α_k represents the importance of G_k(x) in the composite regressor and is calculated as follows:

$${\alpha }_k=\ln \left(\frac{1}{p_k^2}\right)$$

(9)

Step 4: Compute regressor weights:

$$\begin{array}{cc}{\omega }_{k+1}\left(i\right)=\frac{{\omega }_k\left(i\right)}{{\displaystyle \sum _{i=1}^m{\omega }_k\left(i\right)}}\{\begin{array}{c}\begin{array}{cc}{p}_k^2& {\epsilon }_{k,i}>\theta \end{array}\\ \begin{array}{cc}1& {\epsilon }_{k,i}\le \theta \end{array}\end{array}& i=1,2,\cdots ,m\end{array}$$

(10)

Step 5: If k < N, proceed to Step 2.

Step 6: Construct a linear strong regressor, as:

$$G\left(x\right)={\displaystyle \sum _{k=1}^N{\alpha }_k{G}_k\left(x\right)}$$

(11)

The advantages of the AdaBoost algorithm lie in two primary aspects. Firstly, it employs a mechanism that substitutes random training sample selection with the same set of training samples but with varying weight distributions. Secondly, it utilizes a weighted voting mechanism instead of the traditional average voting mechanism, enhancing the accuracy of the algorithm and making it is particularly suitable for solving real-world problems.

3.2. Model Construction

The 28 base variables, along with their corresponding 28 × 7 derived variables mentioned in Section 2.1, serve as the initial inputs. Following feature extraction and KPCA-based dimensionality reduction, nine principal components are identified as the input variables for evolution. The wind turbine active power regression prediction model is trained utilizing these nine evolutionary inputs with the AdaBoost method.

A 2.0 MW wind turbine, which has been operational since its connection to the grid in December 2020, serves as a representative sample. According to the operation records and field observations, it consistently achieves superior power generation levels with high reliability. Using the operating data of the sample wind turbine from January 2021 to December 2021 as the training set and the data from January 2022 to June 2022 as the test set, the prediction results of the active power intercepted during 00:00–6:00 on 1 January 2022 are presented in Figure 5.

The prediction error based on the proposed method is significantly lower than the classic LinearRegression model based on the results across all test set predictions. Moreover, with the R-squared = 0.966 for the AdaBoost regression model and R-squared = 0.929 for the LinearRegression model, the AdaBoost model provides a superior fit to the data. The p-value of the residual F-test is 2.82 × 10⁻³⁰ (<0.05), which proves the statistical significance and practical applicability of the model.

The performance change in wind turbines is further quantified by evaluating the daily power generation performance score K_T, which is computed using the calculation method presented in Equation (12). A higher score signifies the superior generation performance of the wind turbine.

$$K_T=\frac{1}{N}{\sum }_{n=1}^N\left(P_{pre}-P_{real}\right)/P_{real}\times 100\%$$

(12)

The predicted results are processed in two stages, resulting in the following outcomes for power generation performance monitoring in Figure 6. It is noted that the power performance typically fluctuates within a range of ±2%.

4. Application Example Analysis

4.1. Case 1: Blade Damage Fault Diagnosis

A 2.5 MW wind turbine located in Eastern China, which was grid-connected and operational in June 2018, was found to have a blade tip broken on 13 January 2021 by the on-site operation and maintenance engineers. The affected wind turbine is numbered 24# in the wind farm. The operational data of turbine 24# are collected, and the operation data from June 2018 to December 2019 are utilized as the training set. The KPCA-AdaBoost method is applied to develop a regression prediction model for wind turbine active power, which enables us to predict the theoretical active power from January 2020 to January 2021. Subsequently, the prediction results undergo secondary processing, utilizing Equation (12), leading to the derivation of the trend in power generation performance changes, as depicted in Figure 7.

Following the application of a ±5% threshold, the power generation performance index K_T of the wind turbine experienced a decline greater than 5% on 25 November 2020, and, subsequently, another drop exceeding 10% on 17 December 2021. Preliminary observations indicate that the emergence of significant blade damage became apparent on turbine 24# on 25 November 2020. The analysis of this deterioration trajectory suggests that critical cracking damage likely occurred on 17 December 2020, which had a significant impact on the power performance of the wind turbine. The validation of the findings demonstrates that the proposed approach is capable of diagnosing blade damage several weeks prior to the critical event, thereby providing a window of opportunity for proactive intervention to mitigate further degradation.

4.2. Case 2: Power Performance Optimization Assessment

A wind farm equipped with 33 1.5 MW wind turbines interconnected to the grid in 2016 is situated in a mountainous region of southwestern China. The average wind speed and power generation of multiple turbines within the site area gradually decline over time, primarily influenced by terrain and environmental variations. To address this concern, the manufacturer fitted a vortex generator to the blade roots of three underperforming turbines in October 2021, aiming to improve the wind energy conversion efficiency of the blades. According to the results of finite element simulation analysis, the vortex generator could theoretically increase the turbine power output by 4.8%.

The improvement in the power performance of the three inefficient turbines before and after the modification was evaluated based on the KPCA-AdaBoost method trained with operational data from January 2020 to December 2020. The trends in power performance changes of these turbines from October 2021 to November 2021 are illustrated in Figure 8. The evaluation results are presented in

Table 1. Summary table of evaluation results.

Turbine Label	Prior KT	Posterior KT	KT Increase	Uncertainty Error
A28	0.26%	5.23%	4.97%	0.17%
A30	−0.15%	4.45%	4.60%	−0.2%
A32	0.08%	5.46%	5.28%	0.48%

, showing an enhancement in the performance of the three inefficient turbines, ranging from 4.60% to 5.38%. These improvements align closely with the anticipated theoretical improvements, with an uncertainty error in the range of ±0.5%. The results demonstrate that the proposed method accurately quantifies the power performance of wind turbines, with assessment uncertainty significantly lower than that of conventional power curve modeling approaches.

5. Conclusions

In this study, a data-driven modeling method is proposed for wind turbine power performance trends monitoring using SCADA data. The method integrates a KPCA approach for characteristic dimensionality reduction and an AdaBoost enhanced regressor for power prediction, which allows for the incorporation of a sufficient number of inputs, including both basic and derived variables, into the model and somehow enhances the prediction performance. The experimental results indicate that the proposed method outperforms traditional approaches like LinearRegression on the test datasets, as it generates more accurate prediction results with smaller prediction errors. Additionally, a day-oriented deviation indicator is constructed to quantify the performance fluctuations. The results of two industrial cases indicate that the proposed method exhibits outstanding performance in early warnings of blade damage and the quantitative assessment of power performance. In the first case, the proposed method can provide early warning alerts for blade damage faults several weeks in advance. In the second case, it demonstrates higher evaluation accuracy compared to power curve modeling approaches, with an assessment uncertainty within ±0.5%. The proposed method establishes advantages in reducing the uncertainty of power generation performance evaluation through sufficient feature construction, decisive data dimensionality reduction, and the reinforcement of learning algorithm, thereby breaking through the dependence on power curves. In future studies, the methodology will be optimized based on more application scenarios to enable broader industrial applications.