Wind turbine generator system-level operation state and risk monitoring method fusing deep learning and Bayesian neural network
By integrating deep learning and Bayesian neural network, the problem of high cost and poor performance of wind turbine status monitoring is solved, and the accurate status monitoring and risk assessment of wind turbine units is realized, which reduces maintenance costs and improves utilization.
Patent Information
- Application Number
- CN202410902925.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-05
- Publication Date
- 2024-12-03
AI Technical Summary
The operating environment of the wind turbine unit is harsh, the existing status monitoring technology is costly and has limited results, making it difficult to achieve more accurate, smarter and lower cost status monitoring, which affects the unit utilization rate and maintenance costs.
Using a method of fusion deep learning and Bayesian neural network, a status monitoring model for key components of wind turbines is constructed through the TimeMixer network, and a health degree is calculated based on JS divergence, and a Bayesian neural network is used to predict system-level health degree and failure probability. Realize system-level operating status and risk monitoring of wind turbines.
It realizes more accurate and intelligent status monitoring and risk assessment of wind turbines, reduces maintenance costs, improves unit utilization, and has a lower cost.
Smart Images

Figure CN119062522A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of intelligent operation and maintenance of wind turbines, and specifically to a method for monitoring system-level operating status and risks of wind turbines by integrating deep learning and Bayesian neural networks. Background Art
[0002] As a major trend in the development of clean energy, wind power generation technology has been booming in my country in the past decade. With the large-scale construction of wind turbines, how to reduce the operation and maintenance costs of wind turbines and how to further improve their utilization rate has become an urgent problem to be solved. Compared with the rapidly developing wind power industry, my country's wind turbine intelligent operation and maintenance technology is relatively backward. The working environment of wind turbines is harsh, and the operating conditions are complex and changeable. The occurrence or superposition of various factors will cause the units to operate in an unreliable state. Wind turbines are a complex multi-body and multi-domain system. Compared with traditional power generation systems, wind turbines have small single-unit capacity, many wind farm units, and relatively high failure rates. Due to the harsh operating environment and complex and changeable loads, the operating reliability of key components of wind turbines is greatly affected. Therefore, how to monitor the state and conduct risk assessment of the operation process of wind turbines to better improve the utilization rate of units and reduce maintenance costs is of great significance in the wind power industry.
[0003] Wind turbines are complex multi-body and multi-domain systems, and condition monitoring is the most effective and economical way to ensure their safe and reliable operation. For the condition monitoring of wind turbines, most megawatt-class units are currently equipped with a Supervisory Control and Data Acquisition (SCADA) system, which generally has built-in equipment condition monitoring and abnormal alarm modules. However, these built-in functional modules are usually based on some simple rules to achieve condition monitoring, and their effects are poor. In addition to the SCADA system, for the transmission system, some wind turbines are equipped with a condition monitoring system (CMS) using vibration signals, which can achieve more accurate and sensitive condition monitoring. However, the cost of the condition monitoring system is relatively high, and it has not been widely promoted in wind farms at this stage. Therefore, in addition to relying on the existing SCADA and CMS systems, it is still necessary to continue to study new technologies for wind turbine condition monitoring in order to achieve more accurate, smarter and lower-cost condition monitoring. Summary of the invention
[0004] The purpose of the present invention is to provide a system-level operating state and risk monitoring method for wind turbines that integrates deep learning and Bayesian neural networks. The method combines Prognostics and Health Management (PHM) with Probabilistic Risk Assessment (PRA) technology to achieve a system-level operating state and risk monitoring method for wind turbines. First, the key state variables of the wind turbine are selected and their operating data are obtained. The missing values and abnormal values of the acquired operating data are processed and standardized. Secondly, the key components are selected according to expert experience and the failure mechanism of the wind turbine. Based on the operating data of each key component, the state monitoring model of the key components of the wind turbine is constructed using the TimeMixer network. Then, the health of the key components of the wind turbine is constructed based on the JS divergence, and the JS divergence of the output value of the state monitoring model of each key component and the actual value is calculated to obtain the health of each key component, thereby realizing the operating state monitoring of the key components of the wind turbine. Finally, the health of each key component and the state variable that has the greatest impact on the wind turbine are used as root nodes to construct a Bayesian neural network topology. The health of each key component and the operating data of the selected state variables are used as input to train the Bayesian neural network, and the system-level health of the wind turbine is predicted by the trained Bayesian neural network to calculate the upper and lower limits of the failure probability of the wind turbine, thereby achieving continuous estimation of the operating risk of the wind turbine.
[0005] The purpose of the present invention is to provide a method for monitoring the system-level operating status and risk of a wind turbine generator system by integrating deep learning and Bayesian neural network, which is specifically implemented according to the following steps:
[0006] Based on the failure mode and component failure mechanism of a certain wind turbine and integrating expert experience, the key components of the wind turbine (i.e., nacelle, gearbox, main shaft, generator, converter, pitch system and impeller) were selected as the objects of operation status monitoring.
[0007] Obtain real-time operating data of relevant state variables of each key component.
[0008] Preprocess missing values and outliers in the running data.
[0009] The maximum and minimum normalization of state variables of different orders of magnitude in the operating data is performed as follows:
[0010]
[0011] Where i is a state variable or a column in the running data. min and x max is the minimum and maximum value of all values of state variable i.
[0012] A sliding window is used to repartition the normalized running data as input.
[0013] According to expert experience, the key state variables of each key component are selected as output, and the TimeMixer network topology is constructed.
[0014] TimeMixer is a multi-scale time series prediction network that can extract basic information from historical sequences from multiple scales through the Past-Decomposable-Mixing (PDM) block, and then use the historical information to predict future sequences through the Future-Multipredictor-Mixing (FMM) block.
[0015] First, TimeMixer averages the historical observation data into Downsampling to M scales, we get a multi-scale representation of the input sequence χ = {x0, x1, ..., x M},in m∈{0,1,…,M}, C represents the number of variables.
[0016] Then, stacked PDM blocks are used to mix historical information at different scales. Let the input of the PDM block of the lth layer be l-1 , then the PDM process can be expressed as:
[0017] χ l =PDM(χ l-1 ),l={0,1,…,L}
[0018] Where L is the total number of PDM blocks. Indicates channel d model A mixed representation of history.
[0019] Therefore, PDM can gradually aggregate detailed seasonal information from fine scale to coarse scale and use coarse scale prior knowledge to deepen macro trend information.
[0020] Finally, the FMM block is used to integrate the multi-scale historical information extracted by the PDM block to predict the future state of the key state variables of each key component. The FMM process can be expressed as:
[0021]
[0022] in, Represents the predicted future state.
[0023] The running data is divided into training set, validation set and test set, and the mean absolute error (MAE) is used as the evaluation index. The TimeMixer network is iteratively trained on the training set, and the model with the smallest MAE value on the validation set is used as the condition monitoring model, and its accuracy is verified on the test set. Similarly, the condition monitoring models of each key component are obtained respectively.
[0024] JS divergence measures the similarity of two probability distributions and solves the problem of asymmetric KL divergence. The KL divergence that measures the difference between two distributions P and Q is:
[0025]
[0026] Where p(x) and q(x) are the probability density functions of distributions P and Q, respectively.
[0027] Then the JS divergence between the two distributions is:
[0028]
[0029] Among them, M=(P+Q) / 2.
[0030] The health of each key component is constructed based on the JS divergence. The larger the JS divergence, the greater the distribution difference between its true value and the model output value.
[0031] The health index (HI) of key components of wind turbines constructed based on JS divergence is as follows:
[0032]
[0033] Among them, Y pred and Y true are the output value and actual value of the key component condition monitoring model, respectively. b is an adjustable hyperparameter that can be used to set the sensitivity of health.
[0034] By bringing the JS divergence between the output value of the condition monitoring model of each key component and the actual value of the operating data into the health degree, the health degree of the key components of the wind turbine is calculated respectively, and the operating status monitoring of the wind turbine subsystem is realized.
[0035] Bayesian Neural Network (BNN) is a neural network that combines the Bayesian method. Unlike traditional neural networks, the weight parameter of BNN is no longer a fixed value, but a probability distribution.
[0036] The parameters in BNN are initialized with a certain prior distribution p(W). In the process of training with training data, following the Bayesian rule, the posterior distribution p(W|D) is updated in each iteration, as shown in the following formula:
[0037]
[0038] Among them, p(D|W) is the likelihood function, p(D) is the marginal likelihood function, and p(W) is the prior distribution of the parameters.
[0039] After training the BNN, we can get the predicted distribution p(y|x,D) for any input x. The formula is as follows:
[0040] p(y|x,D)=∫p(y|x,W)p(W|D)dW
[0041] Among them, y is the output of the model.
[0042] In the network, due to the high dimensionality of the sampling space, the posterior distribution p(W|D) is difficult to obtain, so variational inference is used to train BNN. The idea of variational inference is to use a set of simple distributions to approximate the posterior distribution and use KL divergence to measure the distance between the two distributions. The formula of KL divergence is as follows:
[0043]
[0044] Where W' is the integral variable. θ (W') is the variational distribution used to approximate the posterior distribution p(W|D).
[0045] The training goal of BNN is to minimize D KL (q θ ||p), since p(W|D) still needs to be calculated, the training is carried out by maximizing the evidence lower bound (ELBO), and the formula is as follows:
[0046]
[0047] The health level of each key component sHI={shi1,shi2,…,shi k}(k represents the total number of key components selected) as the root node of BNN.
[0048] According to expert experience, several state variables with the greatest impact on wind turbines are selected. Statistical analysis is performed on the selected state variables, and state variables with a Spearman correlation coefficient of 0.9 or above are removed as redundant variables, and state variables with low temporal variability and a coefficient of variation of less than 0.05 are removed.
[0049] The state variables after removal are also used as the root nodes of BNN. Then the input of BNN model is R = {r1, r2, ..., r n}(n is the total number of root nodes of BNN).
[0050] The BNN topology is constructed, and the state variables of the root node and the historical values of the health of key components are used as input to train the system-level health prediction model of the wind turbine.
[0051] Use the trained model to predict the system-level health of wind turbines and take the average and standard deviation σ cHI , and calculate the upper and lower limits of the probability of wind turbine failure, the formula is as follows:
[0052]
[0053] Where upper and lower are the upper and lower limits of the probability of failure, respectively. ε is an adjustable hyperparameter used to control the size of the interval between the upper and lower limits.
[0054] By inputting the real-time values of the selected state variables and the health of key components through the root node of the BNN network, the upper and lower limits of the probability of wind turbine failure are obtained, thereby achieving continuous assessment of the operating risks of the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 Flowchart of a wind turbine system-level operating status and risk monitoring method integrating deep learning and Bayesian neural network
[0056] Figure 2 TimeMixer structure diagram
[0057] Figure 3 BNN structure diagram Specific implementation methods
[0058] The present invention is further described below in conjunction with the accompanying drawings, but this should not be used to limit the protection scope of the present application.
[0059] In this example, refer to Figure 1 As shown, the present invention proposes a method for monitoring the system-level operating status and risk of a wind turbine generator system by integrating deep learning and Bayesian neural network, comprising the steps of:
[0060] Based on the failure mode and component failure mechanism of a certain wind turbine and integrating expert experience, the key components of the wind turbine (i.e., nacelle, gearbox, main shaft, generator, converter, pitch system and impeller) were selected as the objects of operation status monitoring.
[0061] Obtain real-time operating data of relevant state variables of each key component.
[0062] Preprocess missing values and outliers in the running data.
[0063] The maximum and minimum normalization of state variables of different orders of magnitude in the operating data is performed as follows:
[0064]
[0065] Where i is a state variable or a column in the running data. min and x max is the minimum and maximum value of all values of state variable i.
[0066] A sliding window is used to repartition the normalized running data as input.
[0067] According to expert experience, the key state variables of each key component are selected as output, and the TimeMixer network topology is constructed.
[0068] TimeMixer is a multi-scale time series prediction network that can extract basic information from historical sequences from multiple scales through the Past-Decomposable-Mixing (PDM) block, and then use the historical information to predict future sequences through the Future-Multipredictor-Mixing (FMM) block.
[0069] First, TimeMixer averages the historical observation data into Downsampling to M scales, we get a multi-scale representation of the input sequence χ = {x0, x1, ..., x M},in m∈{0,1,…,M}, C represents the number of variables.
[0070] Then, stacked PDM blocks are used to mix historical information at different scales. Let the input of the PDM block of the lth layer be l-1 , then the PDM process can be expressed as:
[0071] χ l =PDM(χ l-1 ),l={0,1,…,L}
[0072] Where L is the total number of PDM blocks. Indicates channel d model A mixed representation of history.
[0073] Specifically, for the lth PDM block, the multi-scale time series χ l Break down into seasonal components and Trends section To fuse information from multiple scales, the formula is as follows:
[0074]
[0075] Among them, FeedForward(·) contains two linear layers with GELU activation function for information fusion between channels. S-Mix(·) and T-Mix(·) represent seasonal and trend fusion, respectively.
[0076] In seasonal fusion, a bottom-up fusion method is used to merge the fine-scale time series information at the lower level upward, which can supplement the detailed information for coarse-scale seasonal modeling. That is, for a set of multi-scale seasonal components The bottom-up-mixing layer is used for the mth scale in a residual way to realize the bottom-up seasonal information interaction. The process is expressed as:
[0077]
[0078] Among them, Bottom-Up-Mixing (·) is two linear layers with GELU activation function along the time dimension, and its input dimension is The output dimension is
[0079] In trend fusion, a top-down fusion method is adopted to use coarse-scale macro information to guide fine-scale trend modeling. The residual method is used to adopt the Top-Down-Mixing layer for the Mth scale to realize the top-down trend information interaction. The process is expressed as:
[0080]
[0081] Among them, Top-Down-Mixing (·) is two linear layers with GELU activation function, and its input dimension is The output dimension is
[0082] Therefore, PDM can gradually aggregate detailed seasonal information from fine scale to coarse scale and use coarse scale prior knowledge to deepen macro trend information.
[0083] Finally, the FMM block is used to integrate the multi-scale historical information extracted by the PDM block to predict the future state of the key state variables of each key component. After L PDM blocks, the multi-scale historical information is obtained. Then the FMM block formula is:
[0084]
[0085] in, represents the future predicted value from the mth sequence. Indicates the final FMM output value. m (·) denotes the predictor for the m-th scale sequence, i.e., a single linear layer.
[0086] The operating data is divided into training set, validation set and test set according to the ratio of 6:3:1, and the mean absolute error (MAE) is used as the evaluation index. The TimeMixer network is iteratively trained on the training set, and the model with the smallest MAE value on the validation set is used as the condition monitoring model, and its accuracy is verified on the test set. Similarly, the condition monitoring models of each key component are obtained respectively.
[0087] JS divergence measures the similarity of two probability distributions and solves the problem of asymmetric KL divergence. The KL divergence that measures the difference between two distributions P and Q is:
[0088]
[0089] Where p(x) and q(x) are the probability density functions of distributions P and Q, respectively.
[0090] Then the JS divergence between the two distributions is:
[0091]
[0092] Among them, M=(P+Q) / 2.
[0093] The health of each key component is constructed based on the JS divergence. The larger the JS divergence, the greater the distribution difference between its true value and the model output value.
[0094] The health index (HI) of key components of wind turbines constructed based on JS divergence is as follows:
[0095]
[0096] Among them, Y pred and Y true are the output value and actual value of the key component condition monitoring model, respectively. b is an adjustable hyperparameter that can be used to set the sensitivity of health.
[0097] By bringing the JS divergence between the output value of the condition monitoring model of each key component and the actual value of the operating data into the health degree, the health degree of the key components of the wind turbine is calculated respectively, and the operating status monitoring of the wind turbine subsystem is realized.
[0098] Bayesian Neural Network (BNN) is a neural network that combines the Bayesian method. Unlike traditional neural networks, the weight parameter of BNN is no longer a fixed value, but a probability distribution.
[0099] The parameters in BNN are initialized with a certain prior distribution p(W). In the process of training with training data, following the Bayesian rule, the posterior distribution p(W|D) is updated in each iteration, as shown in the following formula:
[0100]
[0101] Among them, p(D|W) is the likelihood function, p(D) is the marginal likelihood function, and p(W) is the prior distribution of the parameters.
[0102] After training the BNN, we can get the predicted distribution p(y|x,D) for any input x. The formula is as follows:
[0103] p(y|x,D)=∫p(y|x,W)p(W|D)dW
[0104] Among them, y is the output of the model.
[0105] In the network, due to the high dimensionality of the sampling space, the posterior distribution p(W|D) is difficult to obtain, so variational inference is used to train BNN. The idea of variational inference is to use a set of simple distributions to approximate the posterior distribution and use KL divergence to measure the distance between the two distributions. The formula of KL divergence is as follows:
[0106]
[0107] Where W' is the integral variable. θ (W') is the variational distribution used to approximate the posterior distribution p(W|D).
[0108] The training goal of BNN is to minimize D KL (q θ ||p), since p(W|D) still needs to be calculated, the training is carried out by maximizing the evidence lower bound (ELBO), and the formula is as follows:
[0109]
[0110] The health level of each key component sHI={shi1,shi2,…,shi k}(k represents the total number of key components selected) as the root node of BNN.
[0111] According to expert experience, several state variables with the greatest impact on wind turbines are selected. Statistical analysis is performed on the selected state variables, and state variables with a Spearman correlation coefficient of 0.9 or above are removed as redundant variables, and state variables with low temporal variability and a coefficient of variation of less than 0.05 are removed.
[0112] The state variables after removal are also used as the root nodes of BNN. Then the input of BNN is R = {r1, r2, ..., r n}(n is the total number of root nodes of BNN).
[0113] The BNN topology is constructed, and the state variables of the root node and the historical values of the health of key components are used as input to train the system-level health prediction model of the wind turbine.
[0114] Use the trained model to predict the system-level health of wind turbines and take the average and standard deviation σ cHI , and calculate the upper and lower limits of the probability of wind turbine failure, the formula is as follows:
[0115]
[0116] Where upper and lower are the upper and lower limits of the probability of failure, respectively. ε is an adjustable hyperparameter used to control the size of the interval between the upper and lower limits.
[0117] By inputting the real-time values of the selected state variables and the health of key components through the root node of the BNN network, the upper and lower limits of the probability of wind turbine failure are obtained, thereby achieving continuous assessment of the operating risks of the wind turbine.
[0118] The applicant of the present invention has described the specific implementation of the present invention in conjunction with the drawings of the specification. The technicians in the industry should understand that the above specific implementation and description are only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which are within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A wind turbine system-level operating status and risk monitoring method that integrates deep learning and Bayesian neural network. The TimeMixer network is used to build a status monitoring model for key components of wind turbines, and the Bayesian neural network is used to predict the system-level health of wind turbines based on the health of each key component and related state variables, thereby calculating the upper and lower limits of the probability of failure of wind turbines, and realizing continuous estimation of the operating risk of wind turbines. It is characterized in that Follow these steps to implement: Step 1: Based on the failure mode and component failure mechanism of a certain wind turbine, and integrating expert experience, the key components of the wind turbine (i.e., nacelle, gearbox, main shaft, generator, converter, pitch system and impeller) are selected as the operating status monitoring objects. Step 2: Obtain real-time operating data of relevant state variables of each key component. Step 3: Preprocess and standardize the acquired operation data of key components of wind turbines. Step 4: Use the TimeMixer network to build a condition monitoring model for key components of wind turbines. Step 5: Construct the health of key components of wind turbines based on JS divergence. By calculating the JS divergence between the output value of the status monitoring model of each key component and the actual value, the health of each key component is obtained to realize the operating status monitoring of key components of wind turbines. Step 6: Use the health of each key component and the state variables that have the greatest impact on the wind turbine as root nodes to build a Bayesian neural network topology structure to achieve system-level operating status and risk monitoring of the wind turbine. Step 7: Use the health of each key component and the operating data of the selected state variables as input to train the Bayesian neural network, and use the trained Bayesian neural network to predict the system-level health of the wind turbine to calculate the upper and lower limits of the probability of failure of the wind turbine, thereby achieving continuous estimation of the operating risk of the wind turbine.
2. According to the method for monitoring the system-level operation status and risk of wind turbines by integrating deep learning and Bayesian neural network as described in claim 1, It is characterized in that Specifically, in step 1, based on the failure mode and component failure mechanism of a certain wind turbine, and integrating expert experience, key components of the wind turbine (i.e., nacelle, gearbox, main shaft, generator, converter, pitch system and impeller) are selected as the operating status monitoring objects.
3. According to the method for monitoring the system-level operation status and risk of wind turbines by integrating deep learning and Bayesian neural network as described in claim 1, It is characterized in that Specifically, step 2 includes obtaining real-time operating data of state variables related to each key component.
4. According to claim 1, an intelligent monitoring method for induced draft fan operation status based on TimeGAN data enhancement and MSGNet, It is characterized in that Specifically, step 3 is to pre-process and standardize the acquired operation data of key components of the wind turbine. The specific process is as follows: Step 4.1: Preprocess the missing values and outliers of the running data. Step 4.2: Use mean interpolation method to fill in abnormal data and missing data. Step 4.3, mean interpolation method is as follows: Among them, x i 'As the expected leveling target at time t. i-1 is the previous data point of the data to be processed, x i+1 For the next data point of the data to be processed, take the two data points before and after the data to be processed to find the mean for interpolation. Step 4.4: Perform maximum and minimum standardization on the state variables of different orders of magnitude in the operating data. The formula is as follows: Where i is a state variable or a column in the running data. min and x max is the minimum and maximum value of all values of state variable i.
5. According to the method for monitoring the system-level operation status and risk of wind turbines by integrating deep learning and Bayesian neural network as described in claim 1, It is characterized in that Specifically, step 4 uses the TimeMixer network to build a condition monitoring model for key components of a wind turbine. The specific process is as follows: Step 5.1: Use a sliding window to re-partition the standardized running data as input. Step 5.2: TimeMixer averages the historical observation data Downsampling to M scales, we get a multi-scale representation of the input sequence χ = {x 0 ,x 1 ,…,x M },in m∈{0,1,…,M}, C represents the number of variables. Step 5.3: Use stacked PDM blocks to mix historical information at different scales. Let the input of the PDM block at layer l be l-1 , then the PDM process can be expressed as: x l =PDM(x l-1 ),l={0,1,…,L} Where L is the total number of PDM blocks. Indicates channel d model A mixed representation of history. Step 5.4: Use the FMM block to integrate the multi-scale historical information extracted by the PDM block to predict the future state of the key state variables of each key component. The FMM process can be expressed as: in Represents the predicted future state. Step 5.5: Divide the running data into training set, validation set and test set, and use the mean absolute error as the evaluation index. Iteratively train the TimeMixer network on the training set, and obtain the model with the smallest MAE value on the validation set as the condition monitoring model, and verify its accuracy on the test set. Step 5.6, and so on, to obtain the status monitoring model of each key component.
6. According to the method for monitoring the system-level operation status and risk of wind turbines by integrating deep learning and Bayesian neural network as described in claim 1, It is characterized in that Specifically, step 5 constructs the health of key components of the wind turbine generator set based on the JS divergence, calculates the JS divergence between the output value of the state monitoring model of each key component and the actual value, thereby obtaining the health of each key component and realizing the operation state monitoring of the key components of the wind turbine generator set. The specific process is as follows: Step 6.1, the KL divergence that measures the difference between the two distributions P and Q is: where p(x) and q(x) are the probability density functions of distributions P and Q, respectively. Then the JS divergence between the two distributions is: Where M = (P + Q) / 2. Step 6.2: Build the health of each key component based on JS divergence. The larger the JS divergence, the greater the distribution difference between its true value and the model output value. Step 6.3: Based on the JS divergence, the health index (HI) of the key components of the wind turbine is constructed as follows: Among them, Y pred and Y true are the output value and actual value of the key component condition monitoring model, respectively. b is an adjustable hyperparameter that can be used to set the sensitivity of health. Step 6.5: By bringing the JS divergence between the output value of the status monitoring model of each key component and the actual value of the operating data into the health degree, the health degree of the key components of the wind turbine is calculated respectively, and the operating status monitoring of the wind turbine subsystem is realized.
7. According to the method for monitoring the system-level operation status and risk of wind turbines by integrating deep learning and Bayesian neural network as described in claim 1, It is characterized in that Specifically, step 6 uses the health of each key component and the state variable that has the greatest impact on the wind turbine as the root node to construct a Bayesian neural network topology structure to achieve system-level operating status and risk monitoring of the wind turbine. The specific process is as follows: Step 7.1: The parameters in the Bayesian neural network (BNN) are initialized using a certain prior distribution p(W). In the process of training with training data, follow the Bayesian rule and update the posterior distribution p(W|D) in each iteration. The formula is as follows: Where p(D|W) is the likelihood function, p(D) is the marginal likelihood function, and p(W) is the prior distribution of the parameters. After training the BNN, we can get the predicted distribution p(y|x,D) for any input x. The formula is as follows: p(y|x,D)=∫p(y|x,W)p(W|D)dW Among them, y is the output of the model. Step 7.2: Use the idea of variational inference to train BNN, that is, use a set of simple distributions to approximate the posterior distribution, and use KL divergence to measure the distance between the two distributions. The formula of KL divergence is as follows: in, in, W' is the integral variable. θ (W') is the variational distribution used to approximate the posterior distribution p(W|D). The training goal of BNN is to minimize D KL (q θ ||p), since p(W|D) still needs to be calculated, the training is carried out by maximizing the evidence lower bound (ELBO), and the formula is as follows: Step 7.3: Set the health level of each key component sHI = {shi 1 ,shi 2 ,…,shi k }(k represents the total number of key components selected) as the root node of BNN. Step 7.4: Select several state variables with the greatest impact on wind turbines based on expert experience. Perform statistical analysis on the selected state variables, remove state variables with a Spearman correlation coefficient of 0.9 or above as redundant variables, and remove state variables with a coefficient of variation of less than 0.05 and low temporal variability. Step 7.5: The state variable after removal is also used as the root node of BNN. Then the input of BNN model is R = {r 1 ,r 2 ,…,r n }(n is the total number of root nodes of BNN). Step 7.6: Build a BNN topology structure, use the state variables of the root node and the historical values of the health of key components as input, and train the system-level health prediction model of the wind turbine. Step 7.7: Use the trained model to predict the system-level health of the wind turbine and take its mean. and standard deviation σ cHI , and calculate the upper and lower limits of the probability of wind turbine failure, the formula is as follows: Where upper and lower are the upper and lower limits of the probability of failure, respectively. ε is an adjustable hyperparameter used to control the size of the interval between the upper and lower limits. Step 7.8: Input the selected state variables and the real-time values of the health of key components through the root node of the BNN network to obtain the upper and lower limits of the probability of wind turbine failure, thereby realizing the operation risk monitoring of the wind turbine.
Citation Information
Cited By
Predictive maintenance method and system for photovoltaic water heater based on deep learning
CN119315552A
Lithium battery SOC prediction method based on WaveMixer model
CN120468666A
Lithium battery soc prediction method based on wave mixer model
CN120468666B
