Wind generating set health state evaluation method considering operation condition

Abstract

The invention discloses a wind generating set health state evaluation method considering an operation condition. The method comprises the following steps: preprocessing SCADA historical data of a wind generating set; using an mRMR algorithm to select related characteristic parameters of each prediction parameter; dividing operation conditions of the wind turbine generator; a CNN-BiGRU-based multi-parameter prediction model is constructed under each divided working condition; performing working condition division on historical data of the fan in a healthy operation state, dividing characteristic parameter data under each working condition into a training set and a test set, inputting the training set and the test set into a CNN-BiGRU prediction model to obtain a predicted value, and constructing a standard residual set according to a residual between a predicted value and an actual value of the training set; determining a health state grade index; and determining the current health level of the wind turbine generator according to the real-time SCADA data of the wind turbine generator. The evaluation result of the method is more objective and accurate, and the degradation trend of the wind turbine generator can be monitored in advance.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine generator set status assessment methods, and in particular relates to a wind turbine generator set health status assessment method taking operating conditions into consideration. Background Art

[0002] Timely and accurate understanding of the health status of wind turbines is of great significance for optimizing wind turbine control and operation and maintenance strategies, reducing the operation and maintenance costs of wind turbines, and ensuring safe and efficient operation of wind turbines. At present, there are two main methods for evaluating the health status of wind turbines: model-based and data-driven. The model-based health status evaluation method of wind turbines mainly establishes a subsystem health status model of the wind turbine and evaluates the health status of the subsystem based on the model. This method is difficult to implement, and the evaluation result depends on the accuracy of the established model. Moreover, since the wind turbine subsystem model is established, the overall health status evaluation of the unit cannot be achieved; the data-based health status evaluation method of wind turbines is easy to implement and the evaluation result is relatively accurate. The present invention realizes the health status evaluation of wind turbines based on the unit operation data in the wind turbine data acquisition and monitoring system (SCADA). Summary of the invention

[0003] The purpose of the present invention is to provide a method for evaluating the health status of a wind turbine generator set taking operating conditions into consideration. The evaluation result of the method is more objective and accurate, and the deterioration trend of the wind turbine generator set can be monitored in advance.

[0004] The technical solution adopted by the present invention is a method for evaluating the health status of a wind turbine generator set considering operating conditions, which is specifically implemented according to the following steps: Step 1: pre-process the wind turbine SCADA historical data; Step 2: Based on the preprocessed data, the mRMR algorithm is used to select the relevant characteristic parameters of each prediction parameter; Step 3, dividing the wind turbine operating conditions; Step 4: Construct a multi-parameter prediction model based on CNN-BiGRU under various working conditions; Step 5: Divide the historical data of the wind turbine under healthy operation into working conditions according to step 3, divide the characteristic parameter data under each working condition into a training set and a test set, input the data into the CNN-BiGRU prediction model, obtain the predicted value, and then construct a standard residual set based on the residual between the predicted value and the actual value of the training set; Step 6: Determine the health status level indicator; Step 7: Determine the current health level of the wind turbine according to the real-time SCADA data of the wind turbine.

[0005] The present invention is also characterized in that: The data preprocessing process in step 1 is as follows: First, remove the data below the cut-in wind speed and with negative power values. Then, use the quartile method to arrange the wind speed-power and power-wind speed data series in order of size and divide them into four equal parts. Each part accounts for 25% of the total data. The data at the 25%, 50%, and 75% cut-off points are quartiles, which are recorded as the first quartile Q1, the second quartile Q2, and the third quartile Q3, respectively. The difference between the third quartile and the first quartile is the interquartile range IQR:

[0006] will be in ~ Data outside the range are removed.

[0007] In step 2, power, gearbox bearing I temperature, and generator bearing temperature are taken as multi-parameter prediction targets, and the mRMR algorithm is used to screen the candidate parameters, and the candidate parameters with the top six mRMR values ​​are selected as the relevant characteristic parameters of each prediction parameter.

[0008] Step 3 is as follows: Taking wind speed, generator speed and torque as the basis for working condition division, the fuzzy C-means clustering algorithm is used to cluster the data preprocessed in step 1 to realize the division of wind turbine operating conditions and obtain the cluster center of each working condition.

[0009] Step 4 is as follows: A multi-parameter prediction model based on CNN-BiGRU is constructed, and the model is trained based on the sample data of each working condition divided in step 3 to obtain a multi-parameter prediction model under each working condition. At the same time, the sparrow search algorithm is used to optimize the number of hidden nodes, initial learning rate and L2 regularization coefficient of the BiGRU network.

[0010] In step 4, the CNN-BiGRU prediction model includes an input layer, a CNN layer, a BiGRU layer, and an output layer; the convolution kernel size in the prediction model is set to 3, the sliding step size is set to 1, the activation function uses the ReLU function, the pooling layer uses average pooling, the maximum number of training times is set to 100, and the Adam optimization algorithm is used.

[0011] In step 6, the health status level index is determined based on the distribution of the Mahalanobis distance between the residuals of the multi-parameter prediction of the test set in step 5 and the standard residual set.

[0012] Step 6 is as follows: Calculate the Mahalanobis distance between the residual of the multi-parameter prediction value of the test set under the healthy operation state of the wind turbine and the standard residual set, perform sliding average and dynamic weighting on the obtained Mahalanobis distance, and define the health index of the wind turbine set according to the dynamically weighted Mahalanobis distance as follows:

[0013] The wind turbine operating status is divided into three levels: healthy, qualified, and abnormal. The distribution of the Mahalanobis distance between the residual of the multi-parameter prediction result under the healthy operation state of the wind turbine and the standard residual set is The value is determined, and according to the 3σ criterion, the Mahalanobis distance at μ+3σ of the distribution is used as the dividing point between the healthy and qualified grades, and the maximum value of the distribution is used to determine the abnormal state index of the wind turbine.

[0014] Step 7 is as follows: The operating condition of the current wind turbine is identified, and the real-time SCADA data of the current wind turbine is input into the CNN-BiGRU prediction model under the corresponding working condition to obtain the predicted value. Then, the health index of the wind turbine is calculated based on the Mahalanobis distance between the residual of the real-time multi-parameter predicted value and the actual value of the wind turbine to the standard residual set. Finally, the current health level of the wind turbine is determined according to the result of step 6.

[0015] The beneficial effects of the present invention are: The method of the present invention does not need to establish a mathematical model for each subsystem of the wind turbine set, uses easily accessible SCADA data of the wind turbine set, considers the operating conditions of the wind turbine set, and adopts a fuzzy C-means clustering algorithm to divide the operating conditions of the unit. Under various conditions, a multi-parameter CNN-BiGRU prediction model is constructed based on a bidirectional gated recurrent unit neural network, and the CNN-BiGRU network parameters are optimized using a sparrow search algorithm. The residuals between the multi-parameter predicted values ​​and actual values ​​under the healthy operating state of the wind turbine are used to construct a standard residual set. The health assessment index is constructed according to the Mahalanobis distance between the residuals between the real-time predicted values ​​of the multi-parameters of the wind turbine and the actual values ​​to the standard residual set, so as to realize the health status assessment of the wind turbine set. The assessment result is more objective and accurate, and the degradation trend of the wind turbine set can be monitored in advance, so as to provide an early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the method of the present invention; Figure 2 It is the CNN-BiGRU network structure diagram of the present invention; Figure 3 is a wind speed-power scatter plot before data preprocessing in Example 1 of the present invention; Figure 4is a wind speed-power scatter plot after data preprocessing in Example 1 of the present invention; Figure 5 is a result diagram of the working condition division in Example 1 of the present invention; Figure 6 is a curve of the predicted result and the actual monitored value of the power parameter of working condition 1 in embodiment 2 of the present invention; Figure 7 is a curve of the predicted result and the actual monitored value of the power parameter of the second working condition in Example 2 of the present invention; Figure 8 is a curve of predicted results and actual monitored values ​​of three power parameters under working condition in Example 2 of the present invention; Fig. 9 is a curve of the predicted results and actual monitored values ​​of the four power parameters under the working condition in Example 2 of the present invention; Fig.10 is the Mahalanobis distance between the real-time residual and the standard residual set in Embodiment 2 of the present invention; Fig.11 It is the unit health indicator in Example 2 of the present invention. DETAILED DESCRIPTION

[0017] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] The present invention considers the health status assessment method of the wind turbine generator set under the operating conditions, such as Figure 1 As shown, the specific implementation steps are as follows: Step 1: Preprocess the data stored in the wind turbine generator set data acquisition and supervisory control system (SCADA).

[0019] First, the data with wind speeds below the cut-in speed and negative power values ​​are eliminated. Then, the remaining historical data are processed using the quartile method to eliminate abnormal values. The specific process is as follows: The quartile method is used to arrange the wind speed-power (VP) and power-wind speed (PV) data series in order of size and divide them into four equal parts. Each part accounts for 25% of the total data. The data at the 25%, 50%, and 75% cutoff points are called quartiles, which are recorded as the first quartile (Q1), the second quartile (Q2), and the third quartile (Q3), respectively. The gap between the third quartile and the first quartile is also called the interquartile range (IQR): (1) In statistics, it is considered ~ Data outside the range are outliers, so the data outside the above range are eliminated to achieve data preprocessing.

[0020] Step 2: Select 16 parameters such as wind speed, wind turbine speed, generator torque, etc. from the wind turbine SCADA data preprocessed in step 1 as candidate parameters, take power, gearbox bearing I temperature, and generator bearing temperature as multi-parameter prediction targets, and use the maximum relevance and minimum redundancy (Max-Relevance and Min-Redundancy, mRMR) algorithm to screen the 16 candidate parameters, and select the candidate parameters with the top six mRMR values ​​as the relevant characteristic parameters of each prediction parameter.

[0021] The calculation process of mRMR value is: First, the mutual information between each candidate parameter and each predicted parameter is calculated according to formula (2): (2) In formula (2), , , are random variables , The marginal probability density and joint probability density of .

[0022] Then, the maximum correlation index between the feature subset and each prediction parameter is calculated: (3) In formula (3), is a feature subset, is the target variable, For the Features, is the number of features.

[0023] Calculate the minimum redundancy index to minimize the correlation between the feature variables in the selected feature set: (4) Finally, considering the maximum correlation index and the minimum redundancy index criteria, the operator is defined , calculate the mRMR value as: (5) Step 3: Taking wind speed, generator speed and torque as the basis for operating condition division, the fuzzy C-means (FCM) clustering algorithm is used to cluster the data preprocessed in step 1 to realize the division of wind turbine operating conditions and obtain the cluster center of each operating condition.

[0024] FCM achieves soft clustering by introducing membership. Assuming a data set matrix Xfor: (6) In formula (6): n is the total number of samples contained in the sample data set, x i For the i The feature data vector of samples.

[0025] The objective function and constraints of FCM are: (7) (8) In formula (7)~(8): V is the cluster center, U is the membership matrix, c represents the number of cluster categories, n represents the number of samples, u ik For the i The sample k Class membership, m is the fuzziness, For the i Samples to k The distance to the cluster center.

[0026] The general steps of FCM are: 1) Initialize parameters and set the number of clusters c , Fuzziness m , allowable error e and the number of iteration terminations; 2) Initialize the membership matrix U ; 3) According to U j Calculate cluster centers V j ; 4) Calculate the objective function J ; 5) According to V j renew U j+1 , return to step 3), and repeat until the end.

[0027] Step 4: Construct a multi-parameter prediction model based on CNN-BiGRU, and train the model based on the sample data of each working condition divided in step 3 to obtain a multi-parameter prediction model under each working condition. At the same time, the Sparrow Search Algorithm (SSA) is used to optimize the number of hidden nodes, initial learning rate and L2 regularization coefficient of the BiGRU network.

[0028] The CNN-BiGRU prediction model is a network model that combines a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU). Its structure is as follows: Figure 2 As shown in the figure, it includes input layer, CNN layer, BiGRU layer and output layer. The convolution kernel size in the prediction model is set to 3, the sliding step is 1, the activation function uses the ReLU function, the pooling layer uses average pooling, the maximum number of training times is set to 100, and the Adam optimization algorithm is used.

[0029] Step 5: Use step 3 to divide the historical data of the fan under healthy operation into working conditions. According to the characteristic parameters selected in step 2, the characteristic parameter data under each working condition is divided into a training set and a test set, and input into the CNN-BiGRU prediction model trained in step 4 to obtain the predicted value. Then, the standard residual set is constructed according to the residual between the predicted value and the actual value of the training set, as shown in the following formula: (9) In formula (9), is the standard residual set, is the predicted value, is the actual value.

[0030] Step 6: Determine the health status level index based on the distribution of the Mahalanobis distance (MD) between the residuals of the multi-parameter prediction of the test set in step 5 and the standard residual set.

[0031] According to formula (10), the Mahalanobis distance between the residual of the parameter prediction value of the test set and the standard residual set is calculated: (10) In formula (10), for The residual of the multi-parameter forecast at time is the standard residual set, and They are distributed The mean and covariance of .

[0032] In order to eliminate the influence of random errors, the size of the sliding window is set to m, and the sliding average Mahalanobis distance at time t is obtained as: (11) The Mahalanobis distance of the multi-parameter prediction results after sliding average is dynamically weighted, and the weight update formula is: (12) The health index of wind turbines is defined according to the Mahalanobis distance as follows: (13) The operating status of wind turbines is divided into three levels: healthy, qualified, and abnormal, and health indicators are set The distribution of the Mahalanobis distance between the residuals of the multi-parameter prediction under the healthy operation state of the wind turbine and the standard residual set is According to the 3σ criterion, the Mahalanobis distance at μ+3σ of the distribution is used as the dividing point between the healthy and qualified grades, and the maximum value of the distribution is used to determine the abnormal state index of the wind turbine.

[0033] Step 7: Identify the operating conditions of the current wind turbine, input the real-time SCADA data of the current wind turbine into the CNN-BiGRU prediction model under the corresponding working conditions to obtain the predicted value, and then calculate the health index of the wind turbine according to the Mahalanobis distance between the residual of the real-time multi-parameter predicted value and the actual value of the wind turbine to the standard residual set, and finally determine the current health level of the wind turbine according to the result of step 6.

[0034] Embodiment 1: This embodiment takes a 2MW wind turbine as an example and determines the health index based on its SCADA historical data: The wind speed-power scatter plot of the pre-processed data obtained in this embodiment is as follows: Figure 3 As shown, the wind speed-power scatter plot of the preprocessed data is as follows Figure 4 As shown in Figure 2, the results of using the fuzzy C-means clustering algorithm to divide the working conditions are as follows: Figure 5 As shown, it is divided into four working conditions. Tables 1-3 show the mRMR values ​​of different characteristic variables, and the characteristic parameters with the mRMR values ​​in the top six are selected as the input features of the model.

[0035] Table 1 mRMR values ​​of power-related characteristic variables

[0036] Table 2 mRMR values ​​of generator bearing temperature-related characteristic variables

[0037] Table 3 mRMR values ​​of temperature-related characteristic variables of gearbox bearing I

[0038] The results of optimizing the BiGRU network using the sparrow search algorithm are shown in Table 4.

[0039] Table 4 Optimization results

[0040] The operating status of wind turbines is divided into three levels: healthy, qualified, and abnormal, and health indicators are set The distribution of the Mahalanobis distance between the residuals of the multi-parameter prediction under the healthy operation state of the wind turbine and the standard residual set is According to the 3σ criterion, the Mahalanobis distance at μ+3σ of the distribution is used as the dividing point between the healthy and qualified levels, and α=12.74 is obtained. The maximum value of the distribution is used to determine the index of abnormal state alarm of the wind turbine set: The obtained wind turbine health status level indicators are shown in Table 5.

[0041] Table 5 Wind turbine health status index

[0042] Embodiment 2: This embodiment selects 5000 sets of SCADA actual monitoring data of the wind turbine in the healthy operation state of Embodiment 1 as training samples for multi-parameter prediction network training. Figure 6-Figure 9 As shown, they are respectively the power parameter prediction results and actual monitoring value curves of the CNN-BiGRU network of the present invention under various working conditions. It can be seen that the CNN-BiGRU network constructed by the present invention can accurately predict the various parameter results of the wind turbine under various working conditions.

[0043] The historical data of the unit three days before the abnormal shutdown at 7:30 on June 8, 2017 was used for verification. The characteristic parameters selected according to the mRMR algorithm were used as the input of the model to predict the multiple characteristic parameters. The Mahalanobis distance between the real-time residual of the prediction result and the standard residual set and the corresponding health index value were calculated as follows: Fig.10 , Fig.11 As shown, it can be seen that the Mahalanobis distance began to increase at 6:10 on June 7. Fig.11 The wind turbine health index determined by the Mahalanobis distance is shown in the figure. It can be found that the health index has dropped to the interval [0.86, 0.9] several times since 1:40 on June 7. At 12:10 on June 7, the health index dropped below 0.86, indicating that the wind turbine is in an abnormal state until the unit issues a shutdown alarm. The results show that the method of the present invention detects the abnormal state of the wind turbine 19 hours earlier than the SCADA system. Embodiment 3: The health status assessment method of a wind turbine considering operating conditions is implemented in the following steps: Step 1: pre-process the wind turbine SCADA historical data; Step 2: Based on the preprocessed data, the mRMR algorithm is used to select the relevant characteristic parameters of each prediction parameter; Step 3, dividing the wind turbine operating conditions; Step 4: Construct a multi-parameter prediction model based on CNN-BiGRU under various working conditions; Step 5: Divide the historical data of the wind turbine under healthy operation into working conditions according to step 3, divide the characteristic parameter data under each working condition into a training set and a test set, input the data into the CNN-BiGRU prediction model, obtain the predicted value, and then construct a standard residual set based on the residual between the predicted value and the actual value of the training set; Step 6: Determine the health status level indicator; Step 7: Determine the current health level of the wind turbine according to the real-time SCADA data of the wind turbine.

[0044] Embodiment 4: Based on Example 3, the data preprocessing process of step 1 is specifically as follows: First, remove the data below the cut-in wind speed and with negative power values. Then, use the quartile method to arrange the wind speed-power and power-wind speed data series in order of size and divide them into four equal parts. Each part accounts for 25% of the total data. The data at the 25%, 50%, and 75% cut-off points are quartiles, which are recorded as the first quartile Q1, the second quartile Q2, and the third quartile Q3, respectively. The difference between the third quartile and the first quartile is the interquartile range IQR:

[0045] will be in ~ Data outside the range are removed.

[0046] Embodiment 5: On the basis of Example 4, in step 2, power, gearbox bearing I temperature, and generator bearing temperature are used as multi-parameter prediction targets, and the mRMR algorithm is used to screen the candidate parameters, and the candidate parameters with the top six mRMR values ​​are selected as the relevant characteristic parameters of each prediction parameter.

[0047] Embodiment 6: Based on Example 5, step 3 is specifically as follows: Taking wind speed, generator speed and torque as the basis for working condition division, the fuzzy C-means clustering algorithm is used to cluster the data preprocessed in step 1 to realize the division of wind turbine operating conditions and obtain the cluster center of each working condition.

Claims

1. A method for evaluating the health status of a wind turbine generator set considering operating conditions, characterized in that: Follow the steps below to implement it: Step 1: pre-process the wind turbine SCADA historical data; Step 2: Based on the preprocessed data, the mRMR algorithm is used to select the relevant characteristic parameters of each prediction parameter; Step 3, dividing the wind turbine operating conditions; Step 4: Construct a multi-parameter prediction model based on CNN-BiGRU under various working conditions; Step 5: Divide the historical data of the wind turbine under healthy operation into working conditions according to step 3, divide the characteristic parameter data under each working condition into a training set and a test set, input the data into the CNN-BiGRU prediction model, obtain the predicted value, and then construct a standard residual set based on the residual between the predicted value and the actual value of the training set; Step 6: Determine the health status level indicator; Step 7: Determine the current health level of the wind turbine according to the real-time SCADA data of the wind turbine.

2. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: The data preprocessing process in step 1 is as follows: First, remove the data below the cut-in wind speed and with negative power values. Then, use the quartile method to arrange the wind speed-power and power-wind speed data series in order of size and divide them into four equal parts. Each part accounts for 25% of the total data. The data at the 25%, 50%, and 75% cut-off points are quartiles, which are recorded as the first quartile Q1, the second quartile Q2, and the third quartile Q3, respectively. The difference between the third quartile and the first quartile is the interquartile range IQR: will be in ~ Data outside the range are removed.

3. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: In step 2, power, gearbox bearing I temperature, and generator bearing temperature are taken as multi-parameter prediction targets, and the mRMR algorithm is used to screen the candidate parameters, and the candidate parameters with the top six mRMR values ​​are selected as the relevant characteristic parameters of each prediction parameter.

4. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: Step 3 is as follows: Taking wind speed, generator speed and torque as the basis for working condition division, the fuzzy C-means clustering algorithm is used to cluster the data preprocessed in step 1 to realize the division of wind turbine operating conditions and obtain the cluster center of each working condition.

5. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: Step 4 is as follows: A multi-parameter prediction model based on CNN-BiGRU is constructed, and the model is trained based on the sample data of each working condition divided in step 3 to obtain a multi-parameter prediction model under each working condition. At the same time, the sparrow search algorithm is used to optimize the number of hidden nodes, initial learning rate and L2 regularization coefficient of the BiGRU network.

6. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: In step 4, the CNN-BiGRU prediction model includes an input layer, a CNN layer, a BiGRU layer, and an output layer; the convolution kernel size in the prediction model is set to 3, the sliding step size is set to 1, the activation function uses the ReLU function, the pooling layer uses average pooling, the maximum number of training times is set to 100, and the Adam optimization algorithm is used.

7. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: In step 6, the health status level index is determined based on the distribution of the Mahalanobis distance between the residuals of the multi-parameter prediction of the test set in step 5 and the standard residual set.

8. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1 or 7, characterized in that: Step 6 is as follows: Calculate the Mahalanobis distance between the residual of the multi-parameter prediction value of the test set under the healthy operation state of the wind turbine and the standard residual set, perform sliding average and dynamic weighting on the obtained Mahalanobis distance, and define the health index of the wind turbine set according to the dynamically weighted Mahalanobis distance as follows: The wind turbine operating status is divided into three levels: healthy, qualified, and abnormal. The distribution of the Mahalanobis distance between the residual of the multi-parameter prediction result under the healthy operation state of the wind turbine and the standard residual set is The value is determined, and according to the 3σ criterion, the Mahalanobis distance at μ+3σ of the distribution is used as the dividing point between the healthy and qualified grades, and the maximum value of the distribution is used to determine the abnormal state index of the wind turbine.

9. The method for evaluating the health status of a wind turbine generator set considering operating conditions according to claim 1, characterized in that: Step 7 is as follows: The operating condition of the current wind turbine is identified, and the real-time SCADA data of the current wind turbine is input into the CNN-BiGRU prediction model under the corresponding working condition to obtain the predicted value. Then, the health index of the wind turbine is calculated based on the Mahalanobis distance between the residual of the real-time multi-parameter predicted value and the actual value to the standard residual set. Finally, the current health level of the wind turbine is determined according to the result of step 6.

Images