Reliability Assessment of Wind Power Converter Considering SCADA Multistate Parameters Prediction Using FP-Growth, WPT, K-Means and LSTM Network

In order to cooperate the wind farm operators with grasping the operation status of wind power converter, a novel reliability assessment strategy is proposed based on supervisory control and data acquisition (SCADA) multistate parameters prediction of permanent magnet synchronous generator (PMSG) wind turbine. The strategy considers “off-line training, on-line matching and assessment”. The operation reliability of wind power converter is obtained via the analysis and weight computing of confidence level, prediction value and actual value of SCADA multistate parameters. In the “off-line training” part, first, the FP-Growth association algorithm is employed to analyze the confidence levels of SCADA variables to the faults of wind power converter. The variables with high confidence level are defined as SCADA multistate parameters. Afterwards, wavelet packet transform (WPT) and K-means algorithm are employed to decompose, reconstruct, normalize and cluster the time series (off-line data) of multistate parameters under normal operation of wind turbine, to improve the generalization capability of long short term memory (LSTM) prediction model. In the part of “on-line matching and assessment”, the actual value time series (on-line data) of multistate parameters are decomposed and reconstructed by WPT. Each normalized sample is matched to the closest cluster centroid, which is generated in the part of “off-line training”. Then the corresponding LSTM model is conducted to predict based on the each sample. The final prediction value is sum of prediction results of entire samples in clusters. Finally, the reliability of wind power converter is assessed by the proposed strategy. The effectiveness of proposed assessment strategy is verified by the experimental results on a PMSG wind power converter.


I. INTRODUCTION
Power converter is a crucial device of wind energy conversion system (WECS) in a wind turbine [1]- [3].It is utilized to control the speed and torque of the generator, simultaneously, apply the active and reactive power to the grid in the modern The associate editor coordinating the review of this manuscript and approving it for publication was Ton Do .
wind turbine system [4]. Over the past decade, a number of investigations and surveys have revealed that as the one of core devices of a wind turbine, the power converter is with the high failure rate, the long downtime, and the low reliability [5]. These have not only greatly increased the production cost of wind power operators, but also seriously impact the stability of grid and even threaten personal safety [6], [7]. Timely and accurate reliability assessment of wind power VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ converter can remind the operators of necessary maintenance and repair of the power converter, so as to effectively reduce the production cost and provide technical support for improving the reliability of wind power converter [8].
A number of studies have been made concerning reliability assessment of wind power converter [9]- [12]. The main research strategies maybe roughly divided into the following two categories: First, compute the lifetime of one of core devices of power converter to assess the reliability. Second, analyze and compute several variables parameter of WECS to assess the reliability. The former needs to know partial electrical parameters of WECS and few environment parameters. By establishing the thermal cycle model of power converter and according to the known reliability coefficients of electronic components, the lifetime of the power converter can roughly be estimated. The latter needs to collect the data of SCADA variables during the normal and failure operation of the power converter for a long time, from which the variables and data characteristics that can reveal whether the power converter is in the failure state are mined, and then the operation status of the power converter is quantitatively analyzed.
As stated in [13], the lifespans of the wind power converter with low and medium speed PMSG are estimated and compared from the thermal cycling of the IGBT module component. However, the obtained reliability metrics of only machine-side converter are rough ranges under extremely severe restrictions, which lead to weak generalization and comparison capability. Health condition of base-plate solder for IGBT module in wind power converter is assessed by analyzing the case temperature difference in [14], but the effects of ambient temperature on heat transfer efficiency of cooling water and on the case temperature are not taken into account, which give other thermal transference of the case of IGBT module. Moreover, as stated in [15], a method to monitor the health of aluminum electrolytic capacitors for enhance the reliability of the modular multilevel converters is introduced.
With the continuous progress and development of science and technology, the wind turbine is developing towards the direction of large-scale, complexity and cost reduction. More and more sensors are utilized in the wind turbine, which makes the variable parameters reflecting the status of each component of the wind turbine more abundant in the SCADA system, thus providing the possibility to assess the reliability of the power converter more comprehensively and accurately. As stated in [16], the health status of wind turbine is assessed by using recurrent neural networks, which is based on the SCADA data of subsystems, but the only a small number of SCADA variables are focused on, and the association levels with the failures of wind turbine are unclear. In this paper, SCADA variables with higher faults confidence levels are selected as multistate parameters to assess the reliability of wind power converter.
The rest of this paper is organized as follows. Section II presents an overview of proposed strategy. Section III reviews the FP-Growth, WPT (including coefficient series normalization), K-means and LSTM algorithm. Afterwards, reliability assessment model of power converter is proposed. Section IV presents results, and discusses the performance with respect to the several methods considered in the investigation. Finally, Section V summarizes the results and conclusions of this study.

II. THE PROPOSED STRATEGY
Until now, the doubly fed induction generator (DFIG) and PMSG wind turbines have occupied the majority of the global wind power market [17], [18]. Initially, the DFIG equipped with partial scale converter became attractive due to the controllability of active and reactive power. However, with the development of technology, the PMSG equipped with full scale converter has more obvious advantages in low-voltage ride-through capability, higher efficiency and power density [19], which is the main development trend of the wind turbine market in the future. Fig. 1 presents the topological graph of the core devices of a PMSG wind turbine system, which includes blade, gear box, PMSG, full scale power converter, filter and transformer. The electric energy converted from wind energy is finally fed to the grid. In this paper, the reliability of power converter of PMSG wind turbine is assessed. The main contributions of this paper are as follows: 1) FP-Growth algorithm is utilized to extract the high confidence level variables of wind turbine SCADA as the multistate parameters.
2) The prediction model of each state parameter based on WPT, K-means algorithm and LSTM network is established under normal operation of wind turbine.
3) Conduct WPT for the ''on-line'' time series of SCADA multistate parameters when wind turbine is operating. Make the normalized coefficient series samples match to the closest cluster centroids, which are generated in the ''off-line training'' part. Then the corresponding LSTM model is conducted to predict based on the samples. The final prediction value is sum of prediction results of entire samples in clusters.
4) The reliability of wind power converter is assessed by proposed strategy, which is via the analysis and weight computing of confidence level, prediction value and actual value of SCADA multistate parameters.

III. THE PROPOSED RELIABILITY ASSESSMENT MODEL A. FP-GROWTH ASSOCIATION RULE ALGORITHM
To distinguish and mine the variables of high confidence level with power converter faults from numerous wind turbine SCADA variables, FP-Growth association rule algorithm is employed in this paper.

1) FP-GROWTH ALGORITHM
Jiawei Han proposed the FP-Growth algorithm with discovering frequent itemsets using frequent pattern tree (FP-Tree) in 2000 [20]. The FP-Growth method indexes the database for fast computing via FP-Tree. The FP-Tree is constructed as follows. Initially the tree contains as root the null item ∅. Afterwards, each tuple t, X ∈ D, where X = i(t) denotes itemset, D denotes database, insert X into FP-Tree, incrementing the count of all nodes along the path that represents X . If X shares a prefix with some previously inserted transaction, then X will follow the same path until the common prefix. For the remaining items in X , new nodes are create under the common prefix, with counts initialized to 1. The FP-Tree is complete when entire transactions have been inserted. At this point, the FP-Tree serves as an index in lieu of the original database. All frequent itemsets can be mined from the tree directly via the FP-Growth method, whose pseudo code is shown in Algorithm 1.
7 else //process projected FP-Trees for each frequent item i 8 foreach i ∈ R in increasing order of sup(i) do 9 X ← P∪ [1] 10 sup(X ) ← sup(i) //sum of cnt(i) for all nodes labeled i 11 F ← F∪{(X , sup(X ))} 12 Given a collection of frequent itemsets F, when a frequent itemset Z ∈ F, look at all proper subsets X ⊂ Z to compute rules of the form: The rule must be frequent since: Thus, only checking whether the rule confidence level satisfies the minconf threshold. Confidence is computed as follows: If c ≥ minconf , then the rule is a strong rule. Otherwise, B. WAVELET PACKET TRANSFORM AND COEFFICIENT SERIES NORMALIZE 1) WAVELET PACKET TRANSFORM Wavelet packets are a generalization of orthonormal and compactly supported wavelets [21]. Pioneered by Coifman, Meyer and Wickerhauser, methods of wavelet packet transform have been successfully used for data compression. Wavelet only decomposes low-frequency part of a signal. Compared with wavelet decomposition, both low-frequency and high-frequency parts of a signal are decomposed by wavelet packet. The corresponding frequency band is adaptively selected to match the signal spectrum, thus improving the time-frequency resolution. Fig. 2 presents the mission profile of signal S based 3-level wavelet packet decomposition. Where A denotes the low-frequency part of the signal decomposed from the previous level; D denotes the high-frequency part of the signal decomposed from the previous level; Numbers 1-14 denote the node codes of coefficient series.
Scaling function ϕ(x) and wavelet function ψ(x) content two-scale equation: where µ n (x)(n = 2p or n = 2p + 1, p = 0, 1, 2, . . .) in (6) and (7) is defined as wavelet packet of orthonormal VOLUME 8, 2020 scaling function. Equations (6) and (7) could be rewritten as follows: Thus, (8) and (9) are the wavelet packet decomposition of signals. Two coefficient series are obtained after each wavelet packet decomposition, and these two coefficient series are decomposed in the next level respectively, that is, the decomposition process should be carried out for the low-frequency part and the high-frequency part, so that the resolution of each part of the signal frequency band can be improved.
Mallat algorithm could be employed to reconstruct the coefficient series after wavelet packet decomposition: (10) and (11) could be rewritten as follows: Then: Equation (14) is the reconstruction algorithm of wavelet packet, where H * and G * are the dual operators of H and G, respectively. The coefficient series decomposed by wavelet packet could be reconstructed to original signal.

2) COEFFICIENT SERIES NORMALIZE
If the signal time series of length n is decomposed and reconstructed to K levels by wavelet packet, there are 2 K node coefficient series C k , whose length is n in the K -th level, k = 1, 2, . . . , 2 K . Each node coefficient series is normalized to be the sample coefficient series matrix X k . If the length of each sample is m, there are n − m + 1 samples in X k .
where, for instance, X k_1 denotes the first sample, which is

C. K-MEANS ALGORITHM
K-means is a kind of clustering algorithms, which is also an unsupervised learning based on data similarity division. It points the more similar samples into the same cluster. Usually, a distance function d(X , Y ) is defined to denote the similarity between samples X and Y. Minkowski distance, Manhattan distance or Euclidean distance is employed. In this paper, Euclidean distance is chosen as the distance function. The two samples are the more similar, the distance is the much less.
Suppose that there are two sample points P and Q, the corresponding coordinates are: Then, the Euclidean distance between P and Q is expressed: K-means defines a constant k first, where k is the number of clusters. Afterwards, normalizing the cluster centroids. The samples are pointed into the most similar clusters by computing the similarity between each sample and cluster centroid. The steps of K-means algorithm could be expressed as follows: 1) Initialize constant k, and k cluster centroids.
2) Iterate the following process until the cluster centroids would not change. i) Compute the similarity between each sample and cluster centroid. Divide the samples into the most similar clusters. ii) Compute the mean of samples divided into each cluster. The mean is defined as the cluster centroid of each cluster. 3) Obtain the final k cluster centroids and the k clusters including samples.

D. LONG SHORT TERM MEMORY NEURAL NETWORK
Long Short Term Memory (LSTM) neural network is an extension Recurrent neural network (RNN), capable of learning long term dependencies [22]. Remembering information for long periods of time is practically its default behavior, not something it struggles to learn. Unlike traditional RNN, LSTM adds three gates, forget gate, input gate and output gate. An LSTM has three of these gates to protect and control cell states. Forget gate decides whether add or remove the historical information in the hidden layer cell. A 1 represents keep, a 0 represents remove.
Input gate decides whether the input information could be permitted add into the hidden layer cell. A 1 represents could, a 0 represents not. 84458 VOLUME 8, 2020 Output gate decides whether transform the output value in present hidden layer cell into next. A 1 represents output, a 0 represents not.
The forward computation formulas of forget gate f t , input gate i t and output gate o t are expressed as follows: where σ is sigmoid activate function; U f is the connecting weights from output of ex-hidden layer cell to present forget gate; W f is the connecting weights from input to forget gate; b f is the bias of forget gate; U i is the connecting weights from output of ex-hidden layer cell to present input gate; W i is the connecting weights from input to input gate; b i is the bias of input gate; U o is the connecting weights from output of ex-hidden layer cell to present output gate; W o is the connecting weights from input to output gate; b o is the bias of output gate. Fig. 3 is the structure of LSTM cell. In the Fig. 3, C t−1 denotes ex-hidden layer cell state; C t denotes candidate state; C t denotes present hidden layer cell state; tanh denotes hyperbolic tangent function; h t is the output of present hidden cell. Their relations with forget gate, input gate and output gate are expressed by the following formulas, respectively:

E. RELIABILITY ASSESSMENT MODEL
With the increasing capacity of wind power converter, reliable and cost-effective wind energy production is more and more significant. In this part, a novel analysis and weight strategy proposed to assess reliability of wind power converter considering the confidence levels, prediction values and actual values of SCADA multistate parameters. Selecting I state parameters S i (i = 1,2,. . . , I ), with whose confidence levels It should be noted that, the prediction values of multistate parameters are obtained and computed based on the normal operation data of the wind turbine. So, the prediction values is should have been the values under the normal operation. Thus, (25) denotes when the difference between actual value and prediction value of state parameter is larger, the reliability of the wind power converter is lower, conversely, the reliability is higher. Defining R (n) = 0, if R (n) ≤ 0; σ is a tiny real number in case the denominators are zero; W i and B i are weight and confidence level of the ith state parameter, respectively. The higher confidence level of state parameter is, the higher weight of reliability influence is.
The reliability of wind power converter is divided into 5 levels as Fig. 4. Where level 1 indicates reliability is in the range of 1-0.8, which denotes the power converter is in normal operation. Level 2 indicates reliability is in the range of 0.8-0.6, which denotes the power converter is in the failure trend, could prevent the possible faults. Level 3 indicates reliability is in the range of 0.6-0.4, which denotes there are hidden failures generate in the power converter, needed to close down and investigate. Level 4 indicates reliability is in the range of 0.6-0.4, which denotes there are failures generate frequently, should close down the wind turbine and maintain immediately. Level 5 indicates reliability is in the range of 0.2-0, which denotes the power converter is in the serious failure, the wind turbine system would shut down and alarm automatically in case harsh accidents occur.

F. MISSION PROFILE OF PROPOSED STRATEGY
In this paper, the proposed strategy for reliability assessment of wind power converter is composed of ''off-line training'' and ''on-line matching and assessment'', and considering WPT, K-means algorithm, LSTM network, analysis VOLUME 8, 2020 and weight. Fig. 5 is the mission profile of reliability assessment of wind power converter.
The specific steps of reliability assessment strategy are as follows.
Off-line training: 1) The FP-Growth association rule is employed to analyze the confidence levels of SCADA variables about the faults of wind power converter.
2) SCADA variables with confidence levels higher than the threshold Q ref are selected as the multistate parameters to assess the reliability of wind power converter.
3) The time series of multistate parameters under normal operation of wind turbine are implemented by WPT. Coefficient series are procured and normalized. 4) Coefficient series samples are conducted by K-means and divided into several clusters. Each cluster corresponds to a clustering centroid. 5) Every cluster's LSTM prediction model is trained by taking use of coefficient series samples in corresponding cluster.
On-line matching and assessment: 1) The actual time series of multistate parameters in the same time range. are implemented by WPT. Obtained coefficient series are procured and normalized as the same as the step 3) in ''off-line training''.
2) Match the coefficient series samples closed to the clustering centroid and pointing into a cluster. The corresponding LSTM models are utilized to predict. The prediction value of a state parameter is sum of prediction value of each LSTM.
3) The reliability of wind power converter is assessed using (25)-(26) depending on confidence levels, prediction values and actual values.

IV. NUMERICAL EXAMPLES A. DATA DESCRIPTION
To verify the validity of proposed reliability assessment strategy, a 2-MW PMSG wind turbine equipped with full scale converter is used as a case study. The most important parameters are listed in Table 1. Simultaneously, to verify the high precision of proposed prediction strategy, which is WPT, K-means and LSTM (WPT-K-LSTM), other 4 prediction strategies are utilized to compare address the same samples data, including: 1, WPT and LSTM (WPT-LSTM); 2, WPT and ELMAN (WPT-ELMAN); 3, only LSTM; 4, only ELMAN neural network (ELMAN). In this section, firstly, the power converter faults data of several PMSG wind turbines owned by a wind farm operator in 2018 are counted. The FP-Growth algorithm is utilized to analyze the confidence levels of 24 SCADA variables about converter faults, and 8 variables with high confidence levels are selected as SCADA multistate parameters. Then, 2000 hours' time series of SCADA multistate parameters of the wind turbine under normal operation are used as the ''off-line training'' samples, and 5 prediction strategies are employed for training and prediction to reflect the high precision of the proposed prediction strategy in this paper. Afterwards, 300 hours' time series of SCADA multistate parameters of the wind turbine under random operation are used as the ''on-line matching and assessment'' samples, which are contracted to predict by 5 prediction strategies, so as to reflect the generalization of the proposed prediction strategy. Finally, the proposed strategy is employed to address the reliability assessment of power converter. The validity of proposed strategy is verified by the log books of the wind farm.

B. ASSOCIATION ANALYSIS BETWEEN WIND TURBINE SCADA VARIABLES AND CONVERTER FAULTS
The FP-Growth association rule is employed to analyze the confidence levels of 24 SCADA variables about the faults of wind power converter. Table 2 shows the faults monitoring of several PMSG power converter systems of a wind farm operator in 2018.  Table 3 is the wind turbine 24 SCADA variables with their codes. Fig. 6 presents the confidence levels of 24 SCADA variables about converter faults based on FP-Growth algorithm. Where the numbers on the abscissa axis are the codes of SCADA variables in Table 3.
It can be seen from Fig. 6 that, the confidence levels of 24 SCADA variables of wind turbine about converter faults are approximately normal distribution in the range of 0.1-0.9, which presents that some variables have a high association with converter faults, while others have a low association with converter faults. If entire 24 SCADA variables are employed as state parameters in the reliability assessment,  the state parameters with low confidence levels will avianize the weight and effect of the state parameters with high confidence levels, thus reducing the effectiveness of reliability. However, if only individual high confidence variables are considered and other variables are ignored, the assessment strategy may not be sensitive to some faults of converter, thus reducing the generalization ability and accuracy of reliability assessment.
Considering the comprehensive impact of the above factors, this paper introduces threshold Q ref , and selects the variables whose confidence levels are greater than threshold Q ref to form the multistate parameters to assess the reliability of converter, so as to achieve a relative balance in the above-mentioned drawbacks. According to the communication with the wind farm technicians and field verification, this paper sets the threshold Q ref = 0.6, which is also a value close to the ''golden ratio'' of 0.618.
Define the SCADA variables with confidence levels higher than the threshold Q ref as the state parameters, which are: output current of converter, power factor, reactive power, generator spindle speed, temperature of converter module, temperature of converter box, temperature of converter cooling air and temperature of main box. In this paper, the above 8 state parameters are utilized to assess reliability of wind power converter, and their codes and confidence levels are shown in Table 4.  Where ELMAN neural network is a global feed forward local recurrent proposed by Elman in 1990 [23]. It is composed of input layer, hidden layer, undertake layer and output layer. The undertake layer takes the hidden layer state of previous time together with the network input of the current time as the input of the hidden layer through connection memory, which plays a role of state feedback. Considering [short-term], the number of hidden layer neurons of ELMAN model is set to 15, and the Sigmoid activation function is employed. The mathematical expression of ELMAN neural network is as follows: where h jt (k) is the output of hidden layer neurons at t time; x it is the value of time series at t time; w i ij is the connection weight between input layer and hidden layer; w r j is the connection weight between undertake layer and hidden layer; w o j is the connection weight between hidden layer and output layer; y t+1 (k) is the prediction output.

2) THE ANALYSIS OF 5 STRATEGIES' ''OFF-LINE TRAINING'' RESULTS
In the ''off-training'' part, the time series of 2000 hours in the same time of 8 state parameters under the normal operation of the wind turbine are selected as samples, and above 5 prediction strategies are utilized to train them respectively, the corresponding prediction models and test results are obtained.
To compare the performance of the 5 prediction strategies and show the advantages of the proposed strategy, this paper selects 4 calculation prediction error indexes to make statistics and evaluation on the prediction results of the 5 strategies.
1) Root Mean Squared Error (RMSE): 2) Mean Absolute Error (MAE): where in (29)-(32),ŷ t is prediction value; y t is actual value; n is the number of nodes in the time series. RMSE and MAE are related to the scale of dependent variable. They are utilized to compare different prediction results of different strategies in the same time series. The smaller the error is, the stronger the prediction ability of the strategy is. MAPE and TIC are independent of the scale of dependent variable. The values of MAPE and TIC are between 0 and 1. When they are equal to 0, it shows that the strategy is the best fitting. Tables 5-12 are the prediction performances of 5 strategies with each state parameter, respectively. The following conclusions can be analyzed and drawn from Tables 5-12: 1) Compared with other 4 prediction strategies, the proposed strategy WPT-K-LSTM presents the best prediction performance for 8 state parameters, the values of RMSE, MAE, MAPE and TIC are the least, and the fitting capability is the highest. 84462 VOLUME 8, 2020      2) Among the 5 prediction strategies, ELMAN has produced the worst prediction performance. When 8 state parameters are predicted by ELMAN, RMSE, MAE and MAPE show the maximum values in 7 state parameters, with the largest prediction error, while TIC is the largest in 8 state parameters, with the lowest fitting capability.
3) In Table 8, the comparison of prediction performance of reactive power of wind turbine, RMSE, MAE and MAPE of ELMAN prediction results are not the maximum values. The experimental results show that the strategy has generated over-fitting phenomenon. While TIC value is the largest of the 5 prediction strategies, so the ELMAN has the lowest fitting capability.
4) The prediction performance of proposed strategy WPT-K-LSTM is higher than that of WPT-LSTM, which is mainly due to the effects of K-means clustering algorithm in the ''off-line training'' part, that is, clustering the data after WPT, which makes the reconstruction data of the test can use the corresponding LSTM model through ''on-line matching'', thus obtaining the prediction results with higher fitting capability.
5) The prediction performance of WPT-LSTM is higher than that of LSTM, and the prediction performance of WPT-ELAMN is higher than that of ELMAN, which illustrates that WPT can improve the fitting capability of prediction results before test.
6) The prediction performance of WPT-LSTM is higher than that of WPT-ELMAN, and the prediction performance of LSTM is higher than that of ELMAN, which illustrates that LSTM presents the high fitting capability for prediction results.

3) THE ANALYSIS OF 5 STRATEGIES' ''ON-LINE MATCHING AND ASSESSMENT'' RESULTS
In the ''on-line matching and assessment'' part, the time series of 300 hours in the same time of 8 state parameters under recent operation (the converter failed during this period) are selected as samples, above 5 prediction strategies are employed, and obtain the prediction results shown as Fig. 7-14. Finally, a novel analysis and weight strategy is proposed and utilized to assess reliability of wind power converter.
It can be seen from Fig. 7-14       all strategies show greater error and lower fitting capability. According to the inspection of maintenance records, this was caused by some fault of converter at the 35 th hour, which did   not return to normal until the 86 th hour. After the 87 th hour, the error and fitting capability of the 5 prediction strategies to the actual values gradually returned to the original level.

D. 5 PREDICTION STRATEGIES AND ANALYSIS OF PREDICITON RESULTS
Base on the confidence levels, actual values and prediction values generated by proposed strategy WPT-K-LSTM of 300 hours, analysis and weight strategy is employed to assess reliability of wind power turbine. Figures 15 and 16 present the reliability curve of wind power converter and the active power curve of wind turbine in 300 hours, respectively.
The following conclusions can be analyzed and drawn from Fig. 15 and 16: 1) At the 35 th hour, the active power of wind turbine suddenly decreased from 2000kW to about 0kW, and remained at about 0kW until the 86 th hour. According to the analysis of wind power converter reliability curve and Figure 4, the wind power converter has a serious failure during this period.
2) In the following 200 hours, although the wind turbine has been in operation and the active power of the wind turbine  has been kept above 400kW, the reliability of wind power converter has been reduced to 0.6-0 for many times, which indicates that the converter of the wind turbine has serious failure potential. In order to avoid serious accidents, the wind turbine should be shut down immediately and the converter should be maintained to ensure normal and stable operation of the wind power converter.

V. CONCLUSION
The main purpose of this paper is proposing a strategy containing FP-Growth, WPT, K-means, LSTM, analysis and weight to conduct the reliability assessment of PMSG wind power converter. Above all, FP-Growth algorithm is utilized to extract the high confidence level variables of wind turbine SCADA as the multistate parameters. Afterwards, the prediction model of each state parameter based on WPT, K-means algorithm and LSTM is established under normal operation of wind turbine. Conduct WPT and normalization for the ''on-line'' time series of SCADA multistate parameters. The corresponding LSTM models are employed to obtain the prediction values of each state parameters. Finally, the reliability of wind power converter is assessed by proposed strategy, which is via the analysis and weight computing of confidence level, prediction value and actual value of SCADA multistate parameters.
The numerical example results reveal that the proposed strategy has the capability to assess the reliability of the PMSG wind power converter, and are well in agreement with the field operation record. He has been a Professor with the School of Control Science and Engineering, Hebei University of Technology, Tianjin, China, and the School of Electrical Engineering, Hebei University of Science and Technology, Shijiazhuang, China. He has authored five books and more than 130 journal and conference papers, and holds 13 U.S. patents and five computer software copyrights. His current research interests include robotics and complex engineering systems.
Dr. Sun was a recipient of many prestigious national awards from China. He was the Director in many societies and committees in China. He is currently the invited Plenary Speaker and the General Co-Chair of many international conferences. Tangshan SIYUAN GONG received the Ph.D. degree in control theory and control engineering from the Hebei University of Technology, China, in 2011. He is currently working with the Hebei Construction and Investment Group Company, Ltd., after graduation. He has authored one book and eight journal and conference papers and holds six computer software copyrights. His current research interests include fault diagnosis and data mining.