Wind turbine generator yaw system fault early warning method based on GRU-CNN and multi-period data fusion

Abstract

The invention relates to a wind turbine generator yaw system fault early warning method based on GRU-CNN and multi-period data fusion, through GRU and CNN network fusion, comprehensive analysis and calculation of SCADA data and CMS data of a wind turbine generator yaw system are carried out, and the early warning accuracy is improved. The problem of long-time dependence of CMS data is solved through GRU neural network modeling, and a mapping relation with the health state of the yaw system of the wind turbine generator is formed. The SCA DA and CMS data of the yaw system of the wind turbine generator are comprehensively analyzed through the GRU-CNN neural network, the false alarm rate is reduced, and the model accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent fault warning of wind turbines in the wind power generation industry, and in particular to a fault warning method for a yaw system of a wind turbine based on GRU-CNN and multi-period data fusion. Background Art

[0002] With the annual increase in the total installed capacity of wind power, the distribution of wind energy resources is unbalanced, the operating environment of wind turbines is generally poor, the failure rate of wind turbines is high, and the operation and maintenance costs account for a large proportion. Today, a large number of wind turbines have been in operation for many years, and the equipment is aging and entering a high-incidence period of failures. In order to reduce operation and maintenance costs and improve the operational reliability of wind turbines, it is necessary to carry out fault warnings for key components of wind turbines, understand the hidden dangers of wind turbine failures in advance, and reduce or even avoid economic losses. The yaw system is one of the key components of a wind turbine. If a failure occurs, it will not only affect the power generation, but may even cause serious consequences such as equipment damage and tower collapse. It is the focus and difficulty of wind turbine fault warning. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the purpose of the present invention is to provide a wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion, which can effectively solve the problem of wind turbine yaw system fault warning.

[0004] The technical solution provided by the present invention is:

[0005] A wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion includes the following steps:

[0006] Step 1: Establish SCADA data of wind turbine yaw system in different health states and CMS data sample sets for several hours before this state;

[0007] Step 2: Calculate the characteristic values ​​of the CMS data in the time domain and frequency domain at different time periods;

[0008] Step 3: Build a GRU-CNN neural network model;

[0009] Step 4: Train the GRU-CNN model using SCADA data and multi-dimensional CMS data feature values;

[0010] Step 5: Use the trained GRU-CNN neural network model to perform fault warning of the wind turbine yaw system.

[0011] Preferably, in step 1, the sample set includes n groups of CMS data related to the yaw system of the wind turbine generator set X={X1, X2, X3, ..., X n}; SCADA data Y = {Y1, Y2, Y3, ..., Y n} and the corresponding yaw system health status Z = {z1, z2, z3, ..., z n}; CMS data is continuous data for more than 120 minutes, and the data interval is 1 second.

[0012] Preferably, the CMS data of a single wind turbine is X i ={x1, x2, x3, ..., x n}, X i Including continuous yaw pressure, continuous yaw noise, continuous yaw motor current, continuous yaw motor voltage, continuous yaw bearing vibration, continuous yaw bearing torque, x1, x2, x3, ..., x n The data interval is 1 second; SCADA data of the yaw system of a single wind turbine i ={y1, y2, y3, ..., y n},Y i Including the start and stop status of the yaw motor, the start and stop status of the unwinding mechanism, the yaw position deviation, the yaw stop position, the yaw speed, the yaw hydraulic brake value, the hydraulic oil level, the yaw angle, the number of yaw actions, and the hydraulic system pressure; the health status of the yaw system includes: healthy, wind error failure, power failure, transmission failure, yaw drive hydraulic failure, and yaw brake failure.

[0013] Preferably, in step 2, 14 characteristic values ​​of each CMS data are calculated within 120 minutes before the corresponding healthy or fault time, with a time interval of 1 minute, for a total of 120 groups, and the characteristic values ​​are:

[0014] Mean

[0015] variance

[0016] Standard Deviation

[0017] Square root amplitude

[0018] RMS value

[0019] Peak value f6(x t )=max(|x tn |);

[0020] Maximum value f7(x t )=max(x tn );

[0021] Minimum value f8(x t )=min(x tn );

[0022] Waveform indicator f9(x t )=f5(x t ) / f1(x t );

[0023] Peak index f 10 (x t )=f6(x t ) / f5(x t );

[0024] Pulse Index f 11 (x t )=f6(x t ) / f1(x t );

[0025] Margin index f 12 (x t )=f6(x t ) / f5(x t );

[0026] Skewness

[0027] Kurtosis

[0028] In the above formula, x t is any CMS data value, N is the total amount of data in a 1-minute time interval;

[0029] The CMS eigenvalue sample set F = {F1, F2, F3, ..., F n}, divide n groups of sample data into training sets {F train , Y train , Z train} and the test set {F test , Y test , Z test}.

[0030] Preferably, in step 3, the GRU-CNN neural network model includes a GRU layer, a CNN layer, a BN layer, a Flatten layer, a Dropout layer, a splicing layer, a pooling layer, and a fully connected layer;

[0031] GRU layer activation function tanh: a is the parameter passed to the function;

[0032] The activation function of the CNN layer is ReLu: ReLu(a)=max(0,a);

[0033] Multi-classification calculation is performed through the softmax function:

[0034] In the above formula, a is the input parameter of the activation function after the neural network layer calculation;

[0035] The CMS feature value data is calculated by the GRU layer, the Dropout layer, and the fully connected layer, and then spliced ​​with the SCADA data by the splicing layer. After multiple CNN layer convolution calculations, the yaw system health status is output through classification by the Softmax function.

[0036] Preferably, in step 4, the accuracy and loss value are used to evaluate the model performance, and the accuracy and loss value are calculated as follows:

[0037] Accuracy

[0038] Loss value

[0039] Among them, TP is the predicted correct sample; FP is the predicted wrong sample; N is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample. test By comparison, if the prediction accuracy is greater than 0.95, the model is considered qualified.

[0040] Preferably, in step 5, after the model training is completed, real-time SCADA data and 120 groups of CMS characteristic value data at 1 minute intervals within the previous 120 minutes are input, and the model outputs the health status of the wind turbine.

[0041] SCADA data and CMS data are both key data reflecting the operating status of the yaw system of a wind turbine. As a continuously changing value, the time domain and frequency domain feature values ​​of CMS data contain key information of its operating status. The GRU network effectively solves the gradient vanishing and gradient explosion problems existing in the traditional RNN by using a gating mechanism, and is particularly suitable for processing data that is dependent for a long time. The design purpose of GRU is to have high performance while maintaining computational efficiency, and is suitable for a wide range of sequence processing tasks. The CNN network is constructed by imitating the visual perception mechanism of organisms, and can perform supervised learning and unsupervised learning. The convolution kernel parameter sharing in its hidden layer and the sparsity of the inter-layer connection enable the convolutional neural network to grid features with a smaller amount of calculation. Therefore, the present invention proposes a fault warning method for the yaw system of a wind turbine based on GRU-CNN and multi-period data fusion. By fusing the GRU and CNN networks, the SCADA data and CMS data of the yaw system of the wind turbine are comprehensively analyzed and calculated, thereby increasing the accuracy of the warning.

[0042] Compared with the prior art, the present invention has the following advantages:

[0043] 1. The long-term dependency of CMS data is solved through GRU neural network modeling, and a mapping relationship with the health status of the wind turbine yaw system is formed.

[0044] 2. Comprehensively analyze the SCADA and CMS data of the wind turbine yaw system through the GRU-CNN neural network to reduce the false alarm rate and improve the model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic flow chart of a fault warning method for a yaw system of a wind turbine provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples.

[0047] like Figure 1 The embodiment shown is a wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion, and the basic steps include:

[0048] S1: Establish SCADA data of wind turbine yaw system in different health states and CMS data sample sets for several hours before this state;

[0049] S2: Calculate the time domain and frequency domain characteristic values ​​in multiple time periods based on CMS data;

[0050] S3: Build a GRU-CNN neural network model.

[0051] S4: GRU-CNN model is trained using SCADA data and multi-dimensional CMS data feature values;

[0052] S5: Fault warning of wind turbine yaw system is performed through the trained GRU-CNN neural network model.

[0053] Embodiment 1 The following is a specific embodiment of a method for early warning of a yaw system failure of a wind turbine generator set by using a GRU-CNN neural network method to establish a mapping relationship between SCADA data, CMS data and health status of the yaw system of the wind turbine generator set:

[0054] Step 1: Collect SCADA data and CMS data of wind turbine yaw system health, wind error fault, power supply fault, transmission fault, yaw drive hydraulic fault, yaw brake fault and other faults in 7 different health states. Select n (n>10000) groups of CMS data X = {X1, X2, X3, ..., X n}, corresponding SCADA data Y = {Y1, Y2, Y3, ..., Y n} and the corresponding time yaw system health state Z = {z1, z2, z3, ..., z n} form a sample set, with no less than 1000 sets of data for each health status. X1, X2, X3, …, X n Each item includes six types of CMS data of a single wind turbine set from 120 minutes before the healthy state record to the healthy state record time, including continuous yaw pressure, continuous yaw noise, continuous yaw motor current, continuous yaw motor voltage, continuous yaw bearing vibration, and continuous yaw bearing torque, with a data interval of 1 second. n Each item includes the SCADA data of the yaw system of a single wind turbine: yaw motor start and stop status, unwinding mechanism start and stop status, yaw position deviation, yaw stop position, yaw speed, yaw hydraulic brake value, hydraulic oil level, yaw angle, yaw action times, hydraulic system pressure, etc. z1, z2, z3, …, z n The 7 different states are coded in natural order 0, 1, 2, …, 6 for the corresponding yaw system health status.

[0055] Step 2.1: For each group of data in the CMS data X, divide it into 120 time periods with a time interval of 1 minute. In each time period, calculate the mean, variance, standard deviation, root amplitude, root mean square value, peak value, maximum value, minimum value, waveform index, peak index, pulse index, margin index, skewness, and kurtosis of each CMS data, and obtain n groups of 120×14 two-dimensional eigenvalues ​​F={F1, F2, F3, …, F n}.

[0056] Step 2.2: The sample set consisting of n groups of data is divided into training set {F train , Y train , Z train} and the test set {F test , Y test , Z test}.

[0057] Step 3: The GRU-CNN neural network model structure is:

[0058] Layer 1: GRU layer, input shape (120, 14), hidden state dimension 60;

[0059] Layer 2: Dropout layer, dropout = 0.3;

[0060] Layer 3: Flatten layer;

[0061] Layer 4: Concatenate layer, connecting SCADA data with Flatten layer output data;

[0062] Layer 5: CNN layer, number of convolution kernels 8, convolution kernel size 10, stride 5, activation function ReLU, input shape (60, 1);

[0063] Layer 6: BatchNormalization layer;

[0064] Layer 7: MaxPooling1D layer;

[0065] Layer 8: CNN layer, number of convolution kernels 16, convolution kernel size 8, stride 5, activation function ReLU;

[0066] Layer 9: BatchNormalization layer;

[0067] Layer 10: MaxPooling1D layer;

[0068] Layer 11: CNN layer, number of convolution kernels 32, convolution kernel size 4, stride 2, activation function ReLU;

[0069] Layer 12: BatchNormalization layer;

[0070] Layer 13: MaxPooling1D layer;

[0071] Layer 14: Dropout layer, dropout = 0.3;

[0072] Layer 15: Dense layer, number of neurons: 64, activation function: ReLu;

[0073] Layer 15: Dense layer, number of neurons: 7, activation function: Softmax;

[0074] Step 4.1: Model training, training set {F train , Y train , Z train The data is input into the GRU-CNN neural network model for model training. The model is trained for 300 epochs. If the accuracy is higher than 0.97, it meets the requirements. Otherwise, the training is repeated.

[0075] Step 4.2: Model validation, the test set {F test , Y test} The data is input into the trained GRU-CNN neural network model for calculation, and the model output results are consistent with Z test For comparison, if the accuracy is greater than 0.95, the model meets the requirements, otherwise repeat step 4.1.

[0076] Step 5: Input the SCADA data of the yaw system at a certain moment and the feature values ​​of the CMS data of the previous 120 minutes into the verified GRU-CNN model, and the model outputs the health status of the yaw system.

[0077] Example 2 The following is a specific example of using the GRU-CNN model to perform fault warning on the SCADA data and CMS data of the yaw system of a certain type of wind turbine in a wind farm in Huanggang, Hubei:

[0078] Step 1: Collect a large amount of SCADA data and CMS data of this model of wind turbine in seven different health states, including yaw system health, wind error fault, power supply fault, transmission device fault, yaw drive fault, yaw brake fault and other faults.

[0079] Step 2: Calculate the feature value of each minute within 120 minutes based on the CMS data, and randomly divide the data set into a training set and a test set in an 8:2 ratio.

[0080] Step 3: Establish the GRU-CNN neural network model.

[0081] Step 4: The training set is input into the GRU-CNN neural network model for training. After 300 epochs, the model accuracy is 0.9912, and the accuracy of the test set is 0.9860.

[0082] Step 5: Connect the model to the plant-level operation management system of the wind farm to monitor the SCADA data and CMS data of the yaw systems of the 21 wind turbines in the field in real time. The early warning situation from March 1 to April 30, 2024 is shown in Table 1. There were 8 alarms in total, which were verified to be the same as the actual situation.

[0083] Table 1 Early warning of a wind farm model

[0084] Seat number Alarm time Early warning failure Actual situation #6 03 / 01 10:20:24 Transmission failure Gearbox wear and broken teeth #3 03 / 05 09:12:16 Yaw drive failure Back pressure valve pressure is too high #15 03 / 17 11:13:34 Yaw brake failure Friction plate wear #7 03 / 19 12:56:09 Wind error fault Wind vane failure #2 03 / 28 17:45:27 Wind error fault Optocoupler component failure #12 04 / 03 05:13:29 Yaw drive failure Deflection limiter loose #14 04 / 14 09:16:34 Transmission failure Gear wear and inaccurate meshing #16 04 / 29 23:24:05 Transmission failure Gearbox wear and broken teeth

Claims

1. A wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion, characterized in that: The following steps are involved: Step 1: Establish SCADA data of wind turbine yaw system in different health states and CMS data sample sets for several hours before this state; Step 2: Calculate the characteristic values ​​of the CMS data in the time domain and frequency domain at different time periods; Step 3: Build a GRU-CNN neural network model; Step 4: Train the GRU-CNN model using SCADA data and multi-dimensional CMS data feature values; Step 5: Use the trained GRU-CNN neural network model to perform fault warning of the wind turbine yaw system.

2. The wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion according to claim 1 is characterized in that: In step 1, the sample set includes n groups of CMS data related to the yaw system of the wind turbine generator set X = {X1, X2, X3, ..., X n }; SCADA data Y = {Y1, Y2, Y3, ..., Y n } and the corresponding yaw system health state Z = {z1,z2,z3,…,z n }; CMS data is continuous data for more than 120 minutes, and the data interval is 1 second.

3. The wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion according to claim 2 is characterized in that: Single wind turbine CMS data X i ={x1,x2,x3,…,x n }, X i Including continuous yaw pressure, continuous yaw noise, continuous yaw motor current, continuous yaw motor voltage, continuous yaw bearing vibration, continuous yaw bearing torque, x1, x2, x3, …, x n The data interval is 1 second; SCADA data of the yaw system of a single wind turbine i ={y1,y2,y3,…,y n },Y i Including the start and stop status of the yaw motor, the start and stop status of the unwinding mechanism, the yaw position deviation, the yaw stop position, the yaw speed, the yaw hydraulic brake value, the hydraulic oil level, the yaw angle, the number of yaw actions, and the hydraulic system pressure; The health status of the yaw system includes: healthy, wind error fault, power failure, transmission failure, yaw drive hydraulic failure, and yaw brake failure.

4. The wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion according to claim 1 is characterized in that: In step 2, 14 eigenvalues ​​of each CMS data are calculated within 120 minutes before the corresponding healthy or fault time, with a time interval of 1 minute, for a total of 120 groups, and the eigenvalues ​​are: Mean x t is any type of CMS data, N is the number of data in the time interval; variance Standard Deviation Square root amplitude RMS value Peak value f6(x t )=max(|x tn |); Maximum value f7(x t )=max(x tn ); Minimum value f8(x t )=min(x tn ); Waveform indicator f9(x t )=f5(x t ) / f1(x t ); Peak index f 10 (x t )=f6(xt) / f5(x t ); Pulse Index f 11 (x t )=f6(x t ) / f1(x t ); Margin index f 12 (x t )=f6(x t ) / f5(x t ); Skewness Kurtosis In the above formula, x t is any CMS data value, N is the total amount of data in a 1-minute time interval; The CMS eigenvalue sample set F = {F1, F2, F3, ..., F n }, divide n groups of sample data into training sets {F train ,Y train ,Z train } and the test set {F test ,Y test ,Z test }.

5. The wind turbine yaw system fault early warning method based on GRU-CNN and multi-period data fusion according to claim 1 is characterized in that: In step 3, the GRU-CNN neural network model includes a GRU layer, a CNN layer, a BN layer, a Flatten layer, a Dropout layer, a splicing layer, a pooling layer, and a fully connected layer; GRU layer activation function tanh: The activation function of the CNN layer is ReLu: ReLu(a)=max(0,a); Multi-classification calculation is performed through the softmax function: In the above formula, a is the input parameter of the activation function after the neural network layer calculation; The CMS feature value data is calculated by the GRU layer, the Dropout layer, and the fully connected layer, and then spliced ​​with the SCADA data by the splicing layer. After multiple CNN layer convolution calculations, the yaw system health status is output through classification by the Softmax function.

6. The wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion according to claim 1 is characterized in that: In step 4, the accuracy and loss value are used to evaluate the model performance, and the accuracy and loss value are calculated as follows: Accuracy Loss value Among them, TP is the predicted correct sample; FP is the predicted wrong sample; N is the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample. test By comparison, if the prediction accuracy is greater than 0.95, the model is considered qualified.

7. The wind turbine yaw system fault warning method based on GRU-CNN and multi-period data fusion according to claim 1 is characterized in that: In step 5, after the model training is completed, the real-time SCADA data and 120 groups of CMS characteristic value data at 1-minute time intervals within the previous 120 minutes are input, and the model outputs the health status of the wind turbine.

Images