A wind turbine fault diagnosis method considering meteorological factors

Abstract

A kind of wind turbine fault diagnosis method considering meteorological factors, wind farm fault information is collected, and meteorological history record when fault occurs;The mechanism of wind turbine fault is analyzed, the initial feature set is determined, the factors causing wind turbine fault are extracted using ReliefF algorithm, and the optimal feature set is selected;A wind turbine fault diagnosis model based on Bayesian network is established, and the extracted features are added as causes;The wind turbine fault diagnosis model based on Bayesian network is trained using the collected training data set, the prior distribution conforming to the data distribution characteristics is found first, the prior state is updated to posterior state using training data, and the optimal decision is made for fault prediction according to the state model learned continuously.The wind turbine is diagnosed by the present application using Bayesian network combined with meteorological factors, which avoids serious consequences caused by diagnosing wind turbine after fault occurs;The accuracy and speed of prediction and diagnosis are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wind power generation technology, in particular to a wind turbine fault diagnosis method considering meteorological factors. BACKGROUND

[0002] With the proposal of the "double carbon" goal, electricity, as a clean and efficient secondary energy, will play an increasingly important role in supporting social and economic development, serving people's energy needs, and building a clean, low-carbon, safe and efficient energy system. Power supply and demand balance is the basis for safe and stable operation of the power system, and is of great significance to China's energy security. However, because the single machine capacity of wind turbine is large, the overall structure of the unit is complex, and the environment of wind power generation is usually in desert, mountainous area and other natural environment, the running condition is very bad, and there are different influences on the operation environment of the unit under different meteorological conditions, and the probability of failure is also different. Therefore, combined with meteorological factors, it is very important to predict and timely diagnose and eliminate the possible faults of the wind turbine for the healthy operation of the wind turbine and the efficiency of wind power generation.

[0003] Chinese patent application (CN111708798A) discloses "a wind turbine fault diagnosis and processing method and system". The modules such as investigation guidance library and logic diagnosis library are connected with the wind turbine to solve the technical problems of untimely fault response, inaccurate fault positioning and insufficient fault elimination experience; the scheme directly faces the field fault processing business, accurately positions the fault and provides processing guidance scheme, quickly and effectively solves the fault. However, it still can only respond after the fault occurs, and cannot prevent the fault in advance, so it cannot better eliminate the influence of abnormal operation state.

[0004] Bayesian model is a method based on probability theory, which represents the relationship between prior knowledge (evidence, signs) and posterior knowledge (phenomenon, conclusion). It is based on Bayes formula, Bayesian statistical inference and Bayesian network, and the random variables existing in the system can be represented by nodes in the Bayesian network, the nodes are connected by lines to show the directional relationship between them, the prior probability and sample information are used to obtain the posterior probability, and it is mainly used to process random information in uncertain information. Each component and structure in the wind turbine can be represented by a node in the Bayesian network, and the connection lines represent the relationship between each component, providing a new fault diagnosis idea for the wind power generation system. SUMMARY

[0005] Aiming at the problems of fault prediction and timely diagnosis of wind power generation system under complex weather conditions, the application provides a wind turbine fault diagnosis method considering weather factors, which combines weather factors and uses Bayesian network to diagnose wind turbine faults, avoids diagnosing wind turbines after faults occur and causes serious consequences, and improves the accuracy and speed of prediction diagnosis.

[0006] The technical scheme adopted by the application is:

[0007] A wind turbine fault diagnosis method considering weather factors, comprising the following steps:

[0008] Step one: collect the wind farm fault information and weather history records when the fault occurs;

[0009] Step two: analyze the wind turbine fault mechanism, determine the initial feature set, use the ReliefF algorithm to extract features of factors causing wind turbine faults, and select the optimal feature set;

[0010] Step three: establish a wind turbine fault diagnosis model based on Bayesian network, and add the features extracted in step two as causes to the cause layer of the wind turbine fault diagnosis model;

[0011] Step four: train the wind turbine fault diagnosis model based on Bayesian network with the collected training data set, first find the prior distribution that meets the data distribution characteristics, update the prior state to the posterior state with the training data, and repeat this process, and make the optimal decision for fault prediction based on the state model learned continuously.

[0012] Step five: test the wind turbine fault diagnosis model trained in step four with the test set.

[0013] In step one, the wind farm fault information and weather factors when the fault occurs are collected together, and each fault data obtained contains fault causes, fault symptoms, fault phenomena, and weather factors.

[0014] In step two, the weather factors causing faults in the initial feature set of the wind turbine fault include temperature, air pressure, humidity, precipitation and snow, and wind speed factors.

[0015] In step two, the ReliefF algorithm is used to calculate the feature weight, and the highest accuracy classification method is selected to extract the features, as follows:

[0016] The feature weight calculation formula is:

[0017]

[0018] In the formula: W(A) and W'(A) respectively represent the feature weight value after iterative update and before iterative update, j=1, 2,..., k, k is the total number of samples, and the feature A is represented by an n-dimensional array, A={a1, a2,..., a n} respectively, and each sample is a point in an n-dimensional space. j} respectively, and each sample is a point in an n-dimensional space.

[0019] diff(A, R, H j ) represents the difference between the sample R and H j Regarding the difference of the feature A, H j and M j respectively represent the nearest same class sample and different class sample in the training set to the sample R; m represents the sampling times; and K represents the number of nearest neighbor samples.

[0020] The sample feature difference calculation formula is:

[0021]

[0022] Randomly select samples R1 and R2 from the sample data set, and R1[A] and R2[A] represent the sample feature points of the samples R1 and R2 that meet the feature A.

[0023] In step three, a wind turbine fault diagnosis model based on a Bayesian network is established, and the Bayesian model formula is:

[0024]

[0025] Wherein: ω i is a fault alarm information, P(X) is a prior probability of node X; P(ω i |X) is a conditional probability given X, P(X|ω i ) is the probability of X occurring under the condition that ω i occurs, also known as posterior probability;

[0026] The prior information of the model includes switch information, electrical quantity information and protection information, and meteorological factors are integrated into the prior information, and the training set samples ω1, ω2, ω3... are brought into the wind turbine fault diagnosis model based on the Bayesian network to estimate the conditional probability distribution.

[0027] The wind turbine fault diagnosis method considering meteorological factors has the following technical effects:

[0028] 1) The present application uses the signs before the failure to predict the occurrence of the failure, and performs control and elimination in advance, avoiding the serious consequences caused by diagnosing the fan after the failure occurs.

[0029] 2) The present application considers the meteorological factor which has a greater influence on the operation of the wind turbine, improves the accuracy and speed of the prediction diagnosis, and reduces the probability of misdiagnosis.

[0030] 3) The present application uses the correlation between the meteorological factor and the failure to make the wind turbine failure diagnosis model based on the Bayesian network find the possible failure in the case of smaller failure signs, further reduce the impact of the failure, not only diagnose the failure node, but also diagnose the potential failure node which has not occurred, so that the wind turbine failure diagnosis model is suitable for both process diagnosis and prior diagnosis.

[0031] 4) The MF-Bayesian Network model proposed in the present application has a diagnosis accuracy of 94.92%, has certain effectiveness, and can provide a basis for the failure diagnosis of the wind turbine. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 It is a schematic diagram of the basic structure of the Bayesian network.

[0033] Figure 2 It is a Bayesian network diagram of the wind turbine failure diagnosis considering the meteorological factor.

[0034] Figure 3 It is a comparison diagram of the classification accuracy and weight after the weight calculation of the failure characteristics.

[0035] Figure 4 It is a MF-Bayesian Network model diagram.

[0036] Figure 5 It is a schematic diagram of the method flow of the present application. DETAILED DESCRIPTION

[0037] A wind turbine failure diagnosis method considering the meteorological factor, comprising the following steps:

[0038] Step one: Collect the failure information of the wind farm, and the meteorological history record when the failure occurs;

[0039] The wind turbine is a power generation equipment which relies on wind as a natural phenomenon as an energy source, and is affected by natural conditions more than the traditional power generation mode. Therefore, it is of great significance to consider meteorological factors when diagnosing faults of the wind turbine. Before fault diagnosis, all faults occurred since the wind farm is built in a specific area are collected. The history record of meteorological conditions is inquired, and the meteorological factors at the time of each fault occurrence are recorded. The wind farm fault information and the meteorological factors at the time of fault occurrence are collected, and each fault data obtained contains fault causes, fault symptoms, fault phenomena and meteorological factors.

[0040] The fault causes include line aging, air humidity on the day, etc.

[0041] The fault symptoms include abnormal signal fluctuation, abnormal switch action, etc.

[0042] The fault phenomena include component failure, abnormal shutdown, etc.

[0043] Step two: analyze the fault mechanism of the wind turbine, determine the initial feature set, use the ReliefF algorithm to extract the features of the factors causing the wind turbine fault, reduce the feature dimension, and select the optimal feature set;

[0044] The optimal feature set is selected as follows:

[0045] According to the feature weight calculation formula, the weight of each feature is calculated, and the features are arranged according to the weight, the weight and classification accuracy rate curve are drawn respectively, the weight threshold value of the fault feature with the highest classification accuracy rate near the average weight is selected, and all fault features with a weight greater than the threshold value form the optimal feature set. The classification has the highest classification accuracy rate.

[0046] In step two, the causes of the typical fault types of the wind turbine are analyzed, and the typical fault types of the wind turbine are sorted out as shown in Table 1:

[0047] Table 1 Typical fault types of wind turbine

[0048]

[0049] It is found that the failure is related to temperature, air pressure, humidity, precipitation, snowfall, wind speed and other factors. The initial feature set of fan failure determined by this contains all possible meteorological factors that can cause failure. The same type of meteorological and the same type of failure are labeled and classified. After preliminary classification, since not all meteorological factors will have a noticeable impact on fan failure, and too many features will make calculation difficult, feature extraction is performed on all related meteorological factors. Redundant data features will make the model structure complex, which is not conducive to fault recognition and diagnosis, and even reduces the diagnosis accuracy. Therefore, feature extraction is needed to remove redundant features. Feature extraction is the variable selection process, which is an important step to achieve effective wind turbine condition monitoring and fault detection. Common feature selection methods can be roughly divided into filter methods, wrapper methods and embedded methods. Among them, ReliefF(relevant features) is a commonly used filter feature selection method. The failure mechanism of the fan is analyzed, all meteorological factors that can cause fan failure are selected, and the feature weight is calculated using the ReliefF(relevant features) algorithm to select the highest accuracy classification method to extract the features.

[0050] Specifically as follows:

[0051] The feature weight calculation formula is:

[0052]

[0053] In the formula, W(A) and W'(A) represent the feature weight values before and after iteration update respectively. j = 1, 2,..., k, k is the total number of samples. Feature A is represented by an n-dimensional array, A = {a1, a2,..., ak}, each sample is a point in n-dimensional space. Class(R) represents the type of R sample, C is the C sample in the C class sample, M n (C) represents the C sample in the nearest different class sample of R, P(Class(R)) is the proportion of R sample type in the total number of samples, P(C) is the proportion of C sample number in the total number of samples, j

[0054] diff(A, R, H j ) represents the difference between sample R and H j about feature A, H j and M j represent the nearest same sample and different sample of sample R in the training set; m represents the number of samples; K represents the number of nearest neighbors.

[0055] The sample feature difference calculation formula is:

[0056] ​

[0057] Randomly select samples R1 and R2 from the sample data set, and R1[A] and R2[A] represent sample feature points in samples R1 and R2 that meet feature A.

[0058] Step three: Establish a wind turbine fault diagnosis model based on a Bayesian network, and add the features extracted in step two as causes to the cause layer of the wind turbine fault diagnosis model. As shown in Figure 1 , the cause layer represents the cause that directly or indirectly leads to the fault, and is the starting point of evidence transmission.

[0059] In a wind power generation system, each element, such as a line, a transformer, etc., is one-to-one corresponding to a node in the Bayesian network, and can be divided into an element node, a system node, and a load node. The Bayesian network is composed of a directed acyclic graph (DAG) and a conditional probability table (CPT), and according to the logical relationship between elements, a corresponding network is established, which is defined as a cause node, a symptom node, and a fault node. The relationship between nodes is represented by a connection line with an arrow, and the direction of the arrow represents the transmission of evidence, as shown in Figure 1 .

[0060] Research shows that in addition to electrical factors, meteorological factors are also a major factor affecting the normal operation of wind turbines. When a traditional Bayesian model is used to predict faults in a power system, information that can be used usually includes switch information, electrical quantity information, protection information, etc. as prior information. The wind turbine fault model proposed in the application integrates meteorological factors into the prior information of the model. The model construction diagram is shown in Figure 2 .

[0061] A wind turbine fault diagnosis model based on a Bayesian network is established,

[0062] The Bayesian model formula is:

[0063]

[0064] wherein ω i is fault alarm information, P(X) is the prior probability of node X; P(ω i |X) is the conditional probability given X, P(X|ω i ) is the probability of X occurring given that ω i occurs, also known as the posterior probability;

[0065] The prior information of the model includes switch information, electrical quantity information, protection information, and meteorological factors are integrated into the prior information.

[0066] The construction of Bayesian model is divided into three main steps: ①, determine the causal dependence between variables; ②, estimate the prior probability distribution; ③, estimate the conditional probability distribution.

[0067] The main advantage of using Bayesian model is that it can update any state of the node, and can make decisions using the updated probability obtained by belief propagation. It also provides a graphical view of the entire process operation. The meteorological factors are integrated into the prior information to complete step ①, determine the causal relationship between variables, so now let the likelihood function of the fault alarm sample information in the training set D be P(X︱ω i ), let the prior probability P(ω) be P(ω) = N(ω; 0, σ 2 ), where: σ 2 is the variance.

[0068] The training set samples ω1, ω2, ω3... are brought into the established Bayesian network-based wind turbine fault diagnosis model to complete step ③, estimate the conditional probability distribution.

[0069] The conditional probability refers to the probability of the occurrence of an event A under the condition that another event B has occurred. As shown in Figure 1 , each table in Figure 1 is the conditional probability table of the corresponding node affected by other nodes.

[0070] Step four: train the Bayesian network-based wind turbine fault diagnosis model with the collected training data set, first find the prior distribution that meets the data distribution characteristics, update the prior state to the posterior state with the training data, and so on. According to the state model learned continuously, make the optimal decision for fault prediction, as follows:

[0071] Train the Bayesian network-based wind turbine fault diagnosis model with the collected training data set. Take N(0, 1) distribution as the prior distribution of the node, input the training data set into the model, and use the Bayesian formula to update the prior state of each node to the posterior state, estimate the conditional probability of each node to form the conditional probability table to complete the training. When an abnormal running state occurs at a certain node of the trained Bayesian model, the evidence is transmitted to each child node and the posterior probability is updated. The node with the maximum posterior probability is the most likely to occur fault node.

[0072] The prior distribution refers to the probability obtained from past experience and analysis.

[0073] The posterior distribution state is the probability estimate closer to the actual situation obtained by correcting the original prior probability of the node based on new information using the Bayesian formula.

[0074] Step five: test the wind turbine fault diagnosis model trained in step four with the test set.

[0075] Embodiment:

[0076] (1): Select a wind farm in a certain area, collect all the faults that have occurred since the wind farm was built. Record the weather conditions at the time of all past faults. Preprocess the collected data. Each fault information should include fault description and weather factors at the time of fault. For example, Table 2:

[0077] Table 2 Weather factors at the time of fault

[0078]

[0079] (2): Consider that temperature rise will lead to a decrease in equipment pressure level and insulation characteristics; high humidity can reduce equipment insulation performance, leading to current leakage, pollution flashover, etc.; lightning may directly cause equipment to withstand lightning overvoltage, etc. After analysis, the preliminary wind turbine fault characteristics are shown in Table 3.

[0080] Table 3 Wind turbine fault characteristics

[0081]

[0082]

[0083] Calculate the weight of the initial fault characteristics and arrange the weights from large to small. The feature weight and accuracy are shown in Table 4.

[0084] Table 4 Feature weight and accuracy

[0085]

[0086] Express the accuracy and weight as Figure 3 , it is found that the classification accuracy is highest when calculating the 7th feature. With the increase of feature quantity, the classification accuracy decreases sharply at the 8th feature. Finally, the threshold value of feature weight is 0.023.

[0087] (3): Divide the elements into cause nodes, symptom nodes, and fault nodes. Use directed arrows to represent the correlation between nodes. Build a Bayesian network representing the relationship between elements in the wind turbine as shown in Figure 4

[0088] (4): Randomly sample a part of the total data as a training set to train the established Bayesian model. The other data is used as a test set to test the established Bayesian model.

[0089] ​(5): Meteorological factors are integrated into prior information, and the conditional probability distribution is estimated using the training set samples. The Bayesian model is used to predict the probability of failure occurrence. All failure data are converted into classification data, where normal represents normal signals and abnormal represents abnormal signals. The data set used to train the Bayesian model is given in Table 5.

[0090] Table 5 Data set used to train the Bayesian model

[0091]

[0092]

[0093] (6): The accuracy verification found that through k-fold cross-validation, 10,000 groups of predicted data were obtained, of which 8,352 groups of normal data (normal) were correctly predicted, 1,648 groups of fault data (abnormal) were correctly predicted, and 572 groups of data were affected by weather, of which 534 groups were correctly predicted. The final results show that the accuracy rate of the MF-Bayesian Network model for fault prediction is 94.92%.

[0094] (7): The MF-Bayesian Network model is verified by using the receiver operating characteristic curve (ROC), and the results show that the AUC value of the proposed model is 0.9049, which verifies the effectiveness of the proposed model and can provide an effective basis for wind turbine fault diagnosis and risk prediction.

Claims

1. A wind turbine fault diagnosis method considering meteorological factors, characterized in that The method comprises the following steps: Step one: collect the wind farm fault information and the meteorological history record when the fault occurs; Step two: analyze the wind turbine fault mechanism, determine the initial feature set, use the ReliefF algorithm to extract the features of the factors causing the wind turbine fault, and select the optimal feature set; The optimal feature set specifically includes: According to the feature weight calculation formula, the weight of each feature is calculated, and the features are arranged according to the weight, the weight and classification accuracy curves are drawn respectively, the weight threshold value of the fault feature with the highest classification accuracy near the average weight is selected, and all fault features with a weight greater than the threshold value form the optimal feature set; Step three: establish a wind turbine fault diagnosis model based on a Bayesian network, and add the features extracted in step two as reasons to the reason layer of the wind turbine fault diagnosis model; Step four: train the wind turbine fault diagnosis model based on the Bayesian network with the collected training data set, find the prior distribution that meets the data distribution characteristics, update the prior state to the posterior state with the training data, and repeat the process, and make the optimal decision for fault prediction according to the state model learned continuously; In step two, the feature weight is calculated using the ReliefF algorithm, and the feature is extracted using the classification method with the highest accuracy, specifically as follows: The feature weight calculation formula is: ； In the formula: and respectively represent the feature weight value after iterative update and before iterative update; , is the total number of samples; the feature is respectively represented by dimensional array, each sample is a point in dimensional space; represents sample type, is class sample, represents the nearest different class sample in the sample distance, is the proportion of sample type in the total number of samples, is the proportion of the number of the class sample in the total number of samples, ; representative sample and about the feature of the difference, with respectively represent the same class sample and different class sample closest to the sample in the training set; represent the number of sampling times; represent the number of nearest neighbor samples; is the proportion of the number of the class samples to the total number of samples; The sample feature difference calculation formula is: ； Randomly select samples from the sample data set and , and represent sample and sample feature points that meet the characteristics in 2. The wind turbine fault diagnosis method considering meteorological factors according to claim 1, characterized in that: In step one, the wind farm fault information and the meteorological factors when the fault occurs are collected together, and each fault data obtained contains fault causes, fault symptoms, fault phenomena, and meteorological factors. 3.The wind turbine fault diagnosis method considering meteorological factors according to claim 1, wherein: In step two, the meteorological factors that cause the fault in the initial feature set of the wind turbine fault include temperature, air pressure, humidity, precipitation and snow, and wind speed factors.

4. The wind turbine fault diagnosis method considering meteorological factors according to claim 1, characterized in that: In step three, the wind turbine fault diagnosis model based on the Bayesian network is established, and the Bayesian model formula is: ； wherein: is the failure alarm information, is the node is the prior probability; is the conditional probability given is also called the posterior probability; is the likelihood function; The prior information of the model includes switch information, electrical quantity information, protection information, and meteorological factors are integrated into the prior information, and training set samples are brought into the established wind turbine fault diagnosis model based on the Bayesian network to estimate the conditional probability distribution.

5. The wind turbine fault diagnosis method considering meteorological factors according to claim 4, characterized in that: In step three, the construction of the Bayesian model includes three main steps: ① determine the causal dependence relationship between variables; ② estimate the prior probability distribution; ③ estimate the conditional probability distribution; Integrating the meteorological factors into the prior information completes step ① and determines the causal relationship between variables; Let the likelihood function of the fault alarm sample information in the training set be , let the prior probability be , wherein: is the variance; the training set sample is brought into the established wind turbine fault diagnosis model based on the Bayesian network, and step ③ is completed to estimate the conditional probability distribution.

Images