Lightning Damage Detection Method Using Autoencoder: A Case Study on Wind Turbines with Different Blade Damage Patterns

Abstract

There have been numerous reported accidents of lightning strikes damaging wind turbine blades, which poses a serious problem. In certain accidents, the blades that were struck by lightning continued to rotate, resulting in breakage due to centrifugal force. Considering this background, wind turbines situated in Japan have been mandated to be equipped with emergency stop devices. Consequently, upon detection of a lightning strike by the device installed on the wind turbine, the turbine is promptly stopped. In order to restart the wind turbine, it is necessary to verify its soundness by conducting a visual inspection. However, conducting prompt inspections can be difficult due to various factors, including inclement weather. Therefore, this process prolongs the downtime of wind turbines and reduces their availability. In this study, an approach was proposed: a SCADA data analysis method using an autoencoder to assess the soundness of wind turbines without visual inspection. The present method selected wind speed and rotational speed as effective features, employing a sliding window for pre-processing, based on previous studies. Besides, the assessment of a trained autoencoder was conducted through the utilization of the confusion matrix and the receiver operating characteristic curve. It was suggested that the availability of wind turbines could be improved by employing this proposed method to remotely and automatically verify the soundness after lightning detection.

1. Introduction

Countries around the world are working to achieve a low-carbon society, as global environmental issues have become a growing concern. One such effort is the increased use of renewable energy sources, such as wind, solar, and geothermal power generation. Among these, wind power generation has attracted attention as a cost-effective power generation method; therefore, large-scale development, including offshore, is underway around the world [1].

Meanwhile, lightning damage to wind turbines has become a serious problem [2,3,4,5]. Wind turbines are vulnerable to lightning strikes because they are often built in open areas with few tall structures and are increasing in size. Especially in Japan, it has been noted that lightning with a large electrical charge occurs frequently in some areas [6,7,8,9,10]. Consequently, lightning damage in this area is particularly serious. In recent years, there have been numerous reported lightning strikes on wind turbines resulting in fatal damage, including cracks or holes in the blades. A multitude of accidents have also been reported in which the rotating blades have been damaged, with centrifugal forces leading to further severe damage [4,11,12]. According to Reference [13], wind turbines in Japan are subject to shutdowns for an average of more than 50 days during the winter months due to lightning strikes.

In order to prevent lightning damage to wind turbines, a range of countermeasures have been under consideration and implementation, including the installation of receptors, diverter strips, and the grounding design of turbine foundations [14,15,16]. In the case of Japan, the “Interpretation of technical standards for wind power generation facilities” was revised to require the installation of lightning detection systems (LDS) in wind turbines, demonstrating a safer countermeasure to lightning strikes. Consequently, wind turbines are capable of stopping rapidly in the event of a lightning strike [17,18,19].

However, it takes a long time to ascertain the soundness of a wind turbine that has been struck by lightning because it requires a visual inspection by a human. Furthermore, a multitude of cases have been observed in which inclement weather or other factors lead to delays in the inspection process, thereby prolonging the downtime of the wind turbine. This is one of the factors that contribute to the reduction in the availability of wind turbines [20]. For this reason, the development of technology capable of automatically and remotely detecting lightning damage to wind turbines is imperative. This technology is expected to become increasingly important in the context of the global development of large-scale offshore wind farms, a development that has been observed to be accompanied by a decline in accessibility.

A proposed methodology for automatically detecting damage to wind turbine blades remotely involves the utilization of data from the Supervisory Control and Data Acquisition (SCADA) system [21]. The SCADA system is a remote control and monitoring system for wind turbines. The implementation of this technology would enable the assessment of wind turbine soundness without the necessity of visual inspections. Therefore, it is expected that the operation of wind turbines can be rapidly restarted after lightning detection, thereby enhancing the availability. The research using SCADA data has already been conducted on methods for detecting anomalies in wind turbines. It has been demonstrated that anomalies in wind turbines can be detected by using AI-based techniques trained on various operational parameters of the turbines [22,23,24,25,26,27]. However, these methods are intended for the prediction of failures resulting from wind turbine aging. For this reason, they do not effectively deal with sudden failure modes, such as those caused by lightning strikes.

The objective of this research is to validate the hypothesis that a basic autoencoder (AE), which models a healthy wind turbine, will be able to detect lightning damage to wind turbines. This research group is currently engaged in an investigation into SCADA data obtained during lightning strikes on wind turbines [28,29]. Based on previous studies demonstrating that wind speed and rotational speed data obtained from the SCADA system can be used to detect lightning damage, an anomaly detection method using the AE is constructed. The AE is an anomaly detection method that can be applied to nonlinear, complex data structures. It is considered advantageous for detecting anomalies in multiple wind turbines that contain complex data.

The structure of this paper is as follows. Section 2 provides an overview of the accident involving two wind turbines damaged by lightning strikes. Section 3 explains the anomaly detection model using the AE and the method for assessing the accuracy of anomaly detection. Section 4 reports the results of applying this model to the SCADA data of two wind turbines to detect anomalies. Additionally, this section assesses the accuracy of the model in detecting anomalies. Finally, Section 5 summarizes this paper.

2. The Accident Overview and SCADA Data

2.1. SCADA Data

The features for the anomaly detection model were selected from SCADA data of the wind turbines based on previous research [28,29]. This research has shown that lightning-induced faults in wind turbines frequently result in anomalous patterns within the relationship between wind speed and rotational speed. The most recent wind turbines are equipped with devices that collect data at a frequency of once per second, whereas many operators have been observed to retain data that are averaged over a span of one or ten minutes because of the volume of data. In this study, the training of the anomaly detection model was based on 1 min data provided by the operator in the form of a dataset.

2.2. Lightning Damage to Wind Turbines

There have been numerous reports of lightning damage to wind turbines along the coast of the Sea of Japan. The following is a detailed description of the accidents involving the two wind turbines that sustained the most damage. At the time of the accidents, the SCADA data of these wind turbines were recorded.

The following is a detailed report of the accident at wind turbine A. A lightning strike directly on the blade resulted in damage to the turbine. At the time of the accident, turbine A did not have an LDS installed, and thus it continued to operate. Consequently, the damage expanded due to centrifugal force, resulting in blade breakage at its base. The blade breakage was detected by alarms in the blade pitch control system and the wind speed and direction sensor. Figure 1a presents an illustration of wind turbine A after the accident. As can be seen in Figure 1a, the shape of the blades changed significantly, suggesting that there was a high probability of an anomaly in the speed of the wind turbine.

The following is a detailed report of the accident at wind turbine B. Turbine B was struck by lightning on nine separate occasions. Despite sustaining direct strikes by lightning on its blades, the turbine continued to operate. Consequently, the tip receptor located at the extremity of the blade was dislodged by either lightning or centrifugal force, resulting in its dispersion into the surrounding environment. The dispersion of the tip receptor was detected by human visual inspection. The specific lightning strike that resulted in significant damage to the blade has not been identified among the nine recorded strikes. In addition, it has been determined that at the time of the accident, wind turbine B was operating at suppressed maximum power (hereinafter referred to as “suppressed operation”). Figure 1b presents an illustration of the wind turbine B after the accident. According to the appearance shown in Figure 1b, the shape of the blades changed, suggesting that there was a high probability of an anomaly in the speed of the wind turbine.

The SCADA data obtained from the two wind turbines are summarized in

Table 1. SCADA data obtained from the wind turbines.

Wind Turbine	Type of SCADA Data	Data Period
		Training Data	Assessment Data
A	1 min average data	The data after repair for about 32 days (rated operation)	The data at the accident for about 170 min (rated operation)
B	1 min average data	The data before the accident for about 33 days (including rated and suppressed operation)	The data at the accident for about 16 days (suppressed operation)
			The data before the accident for about 1000 min (suppressed operation)

. An anomaly detection model was built using the SCADA data shown in Table 1 to test whether blade damage caused by lightning strikes could be detected. First, normal operation data were used as training data. Second, the data from the lightning strike were used as assessment data, with the data before the lightning strike labeled as normal and the data after the lightning strike labeled as abnormal. In addition, the assessment data for wind turbine B included data from normal suppressed operation before the accident occurred. The purpose of this was to verify that proper anomaly detection was possible using data from the suppressed operation.

3. Anomaly Detection Model

3.1. Training Process and Assessment Process

The training process and assessment process for the anomaly detection model used in this study are summarized in Figure 2. As shown in Figure 2, first, features were selected from the SCADA data and were pre-processed. Second, the AE was trained in the model training process using normal operating SCADA data. Note that 10% of the normal data was utilized for the purpose of validation data. Third, the anomaly score on training data was calculated using the trained AE. This anomaly score consequently determined the threshold for anomaly detection, which was calculated using a percentile-based approach. Fourth, the trained AE detected anomalies in the assessment process. These results were then used to assess the anomaly detection accuracy of the model.

These processes were defined using the Python 3.9 programming language and the Pytorch framework, with the training of the models being carried out using graphics processing units (GPUs).

3.2. Features and Pre-Processing

In this study, the wind speed and blade rotational speed were selected as features for the anomaly detection model based on previous research [28,29]. It has also been shown in previous research that accounting for the temporal characteristics of SCADA data can improve the accuracy of anomaly detection. Accordingly, an attempt was made to extract temporal characteristics using a sliding window, as shown in the following Equation (1):

$${\mathit{x}}_{\mathit{d}}^{\left(\mathit{n}\right)}=\left[{\mathit{x}}_{\mathit{d}}^{\left(n-M_{SW}\right)} {\mathit{x}}_{\mathit{d}}^{\left(n-M_{SW}+1\right)} {\mathit{x}}_{\mathit{d}}^{\left(n-M_{SW}+2\right)} \dots {\mathit{x}}_{\mathit{d}}^{\left(\mathit{n}\right)}\right]$$

(1)

Note that ${\mathit{x}}_{\mathit{d}}^{\left(\mathit{n}\right)}$ represents the nth data point in the dth dimension, and ${\mathit{x}}_{\mathit{d}}^{\left(\mathit{n}\right)}$ is a vector quantity. Figure 3 provides a visual representation of the sliding window process. In this example, a sliding window of M_SW = 3 was employed on the two-dimensional feature x, thereby converting it into a six-dimensional feature. In this study, the input data were 12-dimensional vectors. This was due to the application of a sliding window of M_SW = 6 to the wind speed and rotational speed.

Then, pre-processing was applied to the extracted features. First, the data collected during the turbine’s downtime were excluded from both the training and assessment datasets. This was because the method was intended to detect anomalies in wind turbines while they are in operation. Second, 0–1 normalization was performed using the following Equation (2). The features were then converted so that the maximum value was 1 and the minimum value was 0. This facilitated the training process of the AE:

$$x^{\left(n\right)}\text{′}={\displaystyle \frac{x^{\left(n\right)}-\min \mathit{x}}{\max \mathit{x}-\min \mathit{x}}}$$

(2)

3.3. Autoencoder

The AE is an anomaly detection method that utilizes neural networks [30]. The AE employed in this study was the Basic-AE, as illustrated in Figure 4.

The AE is composed of an encoder and a decoder, as illustrated in Figure 4. The encoder is responsible for receiving N-dimensional input data and compressing it into M-dimensional latent variables. Simultaneously, the decoder is responsible for recovering the original N-dimensional input data from the latent variables. The error between the original input data x and the recovered output data y by the decoder is referred to as the reconstruction error. The parameters of the neural network are tuned to minimize this reconstruction error. In this study, the mean squared error (MSE) was defined as the reconstruction error, as shown in Equation (3).

$$MSE={\displaystyle \frac{1}{N}}{\displaystyle \sum }_{n=0}^{N-1}{\left({\mathit{y}}_{\mathit{n}}-{\mathit{x}}_{\mathit{n}}\right)}^2$$

(3)

In an ideal case, a network can be created that can accurately reconstruct any normal data if the AE is trained exclusively on normal data. It is important to note that this AE has been optimized exclusively for normal data. Consequently, when anomalies were input to the AE, reconstruction was less efficient, resulting in the reconstruction error becoming significantly larger. Therefore, if the reconstruction error was defined as the anomaly score, the more abnormal the data, the greater the anomaly score. The presence of an anomaly can be identified through a comparison of the observed anomaly score with a predefined threshold.

In this study, an AE with a hidden layer node number H of 4 was utilized. As stated in Section 3.2, the number of nodes, N, in the input and output layers was 12. The mini-batch training method was employed to optimize the AE, with training being performed for 1000 epochs with a batch size of 128. The hyperparameters of the AE utilized in this study are outlined in

Table 2. Hyperparameters for the autoencoder.

Hyperparameter	Value/Description
Number of input layer nodes	12
Number of hidden layers	1
Number of hidden layers nodes	4
Activation function	ReLU
Iterative method	SGD
Batch size	128
Learning rate	0.01
Number of epochs	1000

.

3.4. Assessment Method

In this study, the anomaly detection accuracy of the model was assessed using a confusion matrix and a receiver operating characteristic (ROC) curve. In this subsection, we provide a concise overview of the confusion matrix and ROC curves.

A confusion matrix is a table that summarizes the true positives, true negatives, false positives, and false negatives, as demonstrated in Figure 5 [31]. The results of the anomaly detection process could be categorized into four distinct events utilizing the confusion matrix, which could then be subjected to a detailed analysis.

The ROC curve is a graphical representation of the false positive rate–true positive rate characteristic as the threshold is varied [32]. The false positive rate (FPR) and true positive rate (TPR) can be calculated using the following Equations (4) and (5) [32]. As elucidated by the definition, the area to the right of the curve increased as the anomaly detection accuracy increased. The area under the curve (AUC) represents this portion and acts as an indicator of model performance. The visualization of these assessment methods was performed by the utilization of the Python 3.9 programming language and the Matplotlib 3.9.1 plotting library:

$$FPR={\displaystyle \frac{FP}{FP+TN}}$$

(4)

$$TPR={\displaystyle \frac{TP}{TP+FN}}$$

(5)

4. Assessment Result

4.1. Wind Turbine A

The anomaly detection model, which was trained on normal data from wind turbine A, was then employed to analyze the SCADA data during a lightning strike. The threshold value was determined as the 95th percentile of the anomaly score during the training process. The results of the anomaly detection are displayed in Figure 6. Figure 6a,b illustrate the true labels, with the blue plots indicating the data before the lightning strike and the red plots indicating the data after the strike. Figure 6c,d illustrate the results of the anomaly detection, with blue plots indicating the data classified as normal by the model and red plots indicating the data classified as abnormal. Figure 6e illustrates the time characteristics of an anomaly score. In addition, Figure 7 shows the learning curve of this model.

As indicated by Figure 6, the anomaly score increased and exceeded the threshold immediately after the lightning strike, thereby ensuring accurate detection as an anomaly. The anomaly score after a lightning strike increased significantly, whereas the anomaly score before a lightning strike remained stable below the threshold. This finding indicates that stable anomaly detection was possible.

Besides, as demonstrated in Figure 7, the losses of both the training and validation data exhibited a comparable decline with each successive training epoch. This result provided confirmation that the overfitting did not occur.

The results presented in Figure 6 were subjected to assessment using the confusion matrix and the AUC of the ROC curve. The AUC and confusion matrix are displayed in Figure 8. The AUC shown in Figure 8a was 0.995, indicating that anomalies were detected with high accuracy. Furthermore, the number of false negatives was only eight, as indicated by the confusion matrix in Figure 8b. For this reason, it can be concluded that this method has the potential to detect blade anomalies after a lightning strike.

4.2. Wind Turbine B

The anomaly detection model, which was trained on normal data from wind turbine B, was then employed to analyze the SCADA data during lightning strikes. The threshold value was determined as the 99th percentile of the anomaly score during the training process. The results of the anomaly detection are displayed in Figure 9. The configuration presented in Figure 9 is similar to that of Figure 6. Note, however, that the data before the first lightning strike were defined as normal and the data after the 9th strike were defined as abnormal. The data in between were undefined and excluded from the assessment data. This is due to the fact that it was not clear which of the nine lightning strikes in this accident caused the blade damage. Undefined data were plotted in grey. In addition, Figure 10 shows the learning curve of this model.

As indicated by Figure 9, the anomaly score increased and exceeded the threshold immediately after the 9th lightning strike, resulting in its detection as an anomaly. This finding suggests that the method may have potential for detecting blade anomalies after a lightning strike.

Besides, as demonstrated in Figure 10, the losses of both the training and validation data exhibited a comparable decline with each successive training epoch. This result provides confirmation that the overfitting did not occur.

Subsequently, the anomaly detection model was then employed to analyze the SCADA data during normal suppressed operation before the lightning accident. The results of the anomaly detection are displayed in Figure 11. The configuration presented in Figure 11 is similar to that of Figure 6.

As illustrated in Figure 11, the presence of numerous false positives suggested a high probability of misidentification of normal suppressed operation as an anomaly. This phenomenon was attributed to the absence of sufficient data concerning suppressed operations within the training dataset, resulting in deficient training of the model in this regard.

The results presented in Figure 9 and Figure 11 were subjected to assessment using the confusion matrix and the AUC of the ROC curve. The ROC curve and confusion matrix are displayed in Figure 12. The ROC curve for the anomaly detection results during a lightning accident is shown as Verification 1. The ROC curve for the anomaly detection results during a lightning accident and the anomaly detection results during normal suppressed operation is shown as Verification 2. As illustrated in Figure 12a, the AUC for Verification 1, which did not consider suppressed operation, was 0.984. This suggests that anomalies could be detected with high accuracy. By contrast, the AUC for Verification 2, which considered suppressed operation, was 0.627. This finding indicates that the anomaly detection model was incapable of handling suppressed operation. Additionally, Figure 12b indicates the presence of numerous false positives.

These findings imply that the rise in the anomaly score before and after a lightning strike, as depicted in Figure 9, was probably due not only to the damaged blade being identified as an anomaly, but also to the incorrect recognition of the suppressed operation status as an anomaly. In order to accurately detect only blade anomalies caused by lightning strikes, it is imperative to eliminate the influence of suppressed operations. This objective can be achieved by extracting feature values that exhibit robustness across different operating states, such as rated and suppressed operations.

5. Conclusions

In this study, an anomaly detection model was developed using an AE based on SCADA data. The objective of this study was to investigate a method for detecting anomalies in wind turbine blades caused by lightning strikes. Based on the results of previous research, the wind speed and rotational speed, as transformed by the sliding window method, were selected as features and employed for training. To verify the accuracy of the anomaly detection model, data from two actual accidents were analyzed.

The findings of the test indicated a high probability of detecting blade anomalies after lightning strikes in both cases. It has been suggested that if it is possible to obtain approximately one month of normal data from the wind turbine, it would be possible to use this model to detect anomalies in any wind turbines as well. However, it should be noted that this method is not applicable during suppressed operation. This challenge can be overcome by extracting feature values that exhibit robustness across different operating states, such as rated and suppressed operations.

It is believed that this method can be used to detect anomalies of varying severity within the same model, ranging from fatal accidents, such as blade breakage, to minor ones, such as dispersion of tip receptors. This is because the AE accurately captured the relationship between wind speed and rotational speed in a healthy turbine, thereby enabling the detection of damage that affected the blade’s lift coefficient. Therefore, this method is applicable to any damage that significantly changes the lift coefficient of the blade.

In order to enhance the reliability of the model, it is important to increase the number of application examples to other wind turbines. A potentially effective approach to achieving this would be to conduct an experimental simulation of lightning damage to the blades and generate abnormal data for verification purposes. In addition, it is naturally essential to continue the collection and analysis of future accident cases.

The findings from this research can be used to quickly detect anomalies in blades after a lightning strike. This technology can be used to remotely monitor the soundness of the blades, which is expected to allow for rapid restarts and improved availability.