Wind turbine generator state monitoring method based on LSTM-MLSOM

Abstract

The invention discloses a wind turbine generator state monitoring method based on LSTM-MLSOM. The method comprises the following steps: S1, acquiring sensor data from an SCADA system of a wind power plant; s2, dividing data selected by different parts into a training data set and a test data set; s3, inputting the processed data of the different parts into the LSTM corresponding to the different parts to obtain a predicted value and a residual value; s4, inputting the residual value of each component into the SOM model corresponding to the bottom layer in the MLSOM to obtain trained SOM models corresponding to different components; s5, obtaining corresponding MQEi values through the respective SOM models of different parts, and determining respective corresponding health thresholds according to the MQEi values of the parts; s6, combining the MQEi values of different parts at the same moment to generate new data, inputting the new data into a top-layer SOM in the MLSOM to obtain a trained top-layer SOM model and an MQEWT value of the whole machine, and determining a health threshold value of the whole machine through the MQEWT value of the whole machine; and S7, inputting test data into the LSTM-MLSOM model to obtain a monitoring result, judging whether the whole wind turbine generator has a fault according to an early warning strategy, and if the whole wind turbine generator is monitored to have the fault, determining the faulted wind turbine generator component according to the MQEi value of the component and the corresponding health threshold value.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine state monitoring, and in particular relates to a wind turbine state monitoring method based on LSTM-MLSOM. Background Art

[0002] To address the increasingly severe energy crisis and ecological challenges, the development and utilization of wind energy, a renewable energy source, is gaining increasing attention worldwide. Wind turbines, which convert wind energy into electricity, are also rapidly developing, with cumulative installed capacity continuing to grow. Wind turbines are often installed in areas with abundant wind energy but harsh natural conditions, making operation and maintenance difficult and costly. Therefore, developing accurate and reliable wind turbine condition monitoring methods to provide early warning of faults is crucial for minimizing losses, reducing accidents, and improving turbine operational reliability.

[0003] With the development of artificial intelligence, data-driven condition monitoring methods have been widely used in the condition analysis of wind turbines. Existing condition monitoring methods usually use normal behavior modeling methods to monitor the condition of wind turbines, but these methods mainly focus on a key component of the wind turbine, or only conduct condition monitoring research on the entire wind turbine, ignoring the information transmission and condition impact between key components and different levels of the entire wind turbine. Summary of the Invention

[0004] The present invention aims to provide a wind turbine condition monitoring method based on a long short-term memory (LSTM) network and a multilayer self-organizing map (MLSOM) network. The method uses an LSTM algorithm to perform modeling and predictive analysis on data under normal operating conditions of the wind turbine. The residual obtained by subtracting the predicted value from the actual value is used as the input of the bottom-level SOM model in the MLSOM model. Since the supervisory control and data acquisition (SCADA) system does not contain data that individually characterizes components or the entire wind turbine, the top-level SOM in the MLSOM model lacks input data. Therefore, the minimum quantization error (MQE) of each component is used as the input of the top-level SOM model of the MLSOM to construct a normal behavior model of the entire wind turbine, thereby realizing condition monitoring of key components and the entire wind turbine.

[0005] In order to solve the above technical problems, the technical solution provided by the present invention is:

[0006] A wind turbine condition monitoring method based on LSTM-MLSOM, the method comprising the following steps:

[0007] S1. Obtain sensor data of different components of the wind turbine related to the generator active power from the wind farm’s SCADA system;

[0008] S2 uses the historical data of different components of the wind turbine when it is generating electricity normally within the predetermined wind speed range and without any faults as a training dataset (one training dataset for each component), and the online operation data as a test dataset (one test dataset for each component), and preprocesses both datasets. Normal power generation without any faults means that the wind turbine operating data are all within the set range.

[0009] S3. Input the pre-processed training data sets of the different components into the corresponding LSTM models respectively, use the LSTM models to retain the nonlinear relationship between different data in each component, and discard irrelevant noise information in the data, so as to obtain LSTM models trained for different components;

[0010] S4. Input the pre-processed training data sets of different components into the trained LSTM models of different components to obtain the predicted values ​​of different components. Then, subtract the predicted values ​​from the actual values ​​at the same time to obtain the residual values ​​of different components.

[0011] S5. Input the residual values ​​of different components into the SOM model corresponding to the bottom layer in the MLSOM to construct normal behavior models of different components. The normal behavior model refers to a mathematical model that can correctly describe the relationship between various parameters and performance of the device when it meets the design requirements and achieves the expected functions. In this patent, the trained SOM model corresponding to different components is the normal behavior model described above;

[0012] S6. Calculate the MQE of different component training data through the trained SOM models corresponding to different components i value, and through the different components of MQE i The value determines the health threshold of the corresponding component;

[0013] S7, MQE of different components at the same time i The value combination generates new data, which is input into the top SOM in the MLSOM to build the normal behavior model of the whole machine and calculate the MQE of the whole machine. WT value;

[0014] S8. MQE based on the top-level SOM model WT The value is used to determine the health threshold of the whole machine, so as to judge whether the wind turbine is in normal operation;

[0015] S9. Input the test data sets of different components of the wind turbine into the corresponding trained LSTM model, output the predicted value, calculate the residual value, and then input the residual value into the trained SOM model to obtain the MQE of different components. i value, MQE i The values ​​are combined and input into the trained top-level SOM model to obtain the MQE of the wind turbine. WT Value; When the MQE of the wind turbine WT If the value exceeds the health threshold of the whole machine for n times continuously, it is determined that the operation state of the wind turbine is abnormal; then, the MQE calculated by the test data sets of different components is used to calculate the MQE value. i The value is compared with the health threshold of the corresponding component. If the MQE of the component i If the value exceeds the corresponding health threshold value n times continuously, the abnormal component in the wind turbine can be determined and repaired and maintained.

[0016] Furthermore, in step S2, the process of preprocessing the acquired wind turbine data is as follows:

[0017] S2.1. Clear abnormal data in historical data;

[0018] S2.2. To eliminate the dimension effect, the cleaned data is normalized.

[0019] Furthermore, in step S2, abnormal data includes missing value data, shutdown data, power-limited data, data outside the normal operating range of the wind turbine, and outlier data. To eliminate the dimensionality effect, the cleaned data is normalized. Normalization is to convert the original data range to the [0,1] interval. The normalization formula is as follows:

[0020]

[0021] Where x i,j norm Represents the normalized value, x i max and x i min Represents the maximum and minimum values ​​of the i-th feature quantity, x i,j Represents the jth data point of the i-th feature value.

[0022] Furthermore, the step S3 includes the following steps:

[0023] S3.1. Set the length of the dataset input sequence and the time step, where the length of the dataset input sequence is the time window size and the time step is the time interval between samples. Construct multiple data subsets of the same size based on the input sequence length and time step.

[0024] S3.2. Set the input feature dimension, number of hidden layers, number of hidden layer neurons, output layer dimension, training cycle, optimization algorithm, and loss function in the LSTM model.

[0025] S3.3. Select a data subset according to the time series and input it into the LSTM model to output the predicted value;

[0026] S3.4. Calculate a loss function value based on the predicted value, and update the model parameters according to the loss function value;

[0027] S3.5. Select the next data subset according to the time step and execute steps S3.3-S3.4 until the training cycle reaches the set number of times.

[0028] Furthermore, the loss function L used by the LSTM model MSE for

[0029]

[0030] Where m represents the total number of data values, y i represents the true value, Represents the predicted value.

[0031] Furthermore, the step S5 includes the following steps:

[0032] S5.1. Set the number of neurons, learning rate, neighborhood radius, and number of iterations in the competitive layer of the SOM network.

[0033] S5.2. Randomly select a data and input it into the SOM network;

[0034] S5.3. Calculate the Euclidean distance between each neuron and the input data, and find the neuron with the smallest distance to the input data. This neuron is called the winning neuron.

[0035] S5.4. A neighborhood is determined with the winning neuron as the center, based on the set neighborhood radius. All neurons in the neighborhood are updated according to the update formula. The completion of the update indicates that one training is complete.

[0036] S5.5. Repeat S5.1-S5.4 until the training times reach the preset times and the training ends.

[0037] Furthermore, the Euclidean distance calculation formula is:

[0038]

[0039] Where d represents the Euclidean distance, x i Represents the i-th dimension value of x, y i Represents the i-th dimension value of y.

[0040] Furthermore, the update formula w used by the SOM network model in S5.4 is j (t+1) is:

[0041] w j (t+1)=w j (t)+η(t)h i,j (t)(X i -w j (t)) (4)

[0042] Where w j (t+1) represents the updated weight, w j (t) represents the unupdated weight, η(t) represents the learning rate, and h i,j (t) is the nearest neighbor function, X i is the input vector.

[0043] Furthermore, in step S6, the calculation formula of the MQE value is as follows:

[0044] MQE=||X i -W BMU || (5)

[0045] Where, X i is the input vector, W BMU is the winning neuron weight vector, ||X i -W BMU || means calculating the Euclidean distance between the input vector and the winning neuron weight.

[0046] Furthermore, in step S6 or S8, the health threshold is determined as follows:

[0047] S6.1. Construct the calculated MQE values ​​into a data set;

[0048] S6.2. Set a sliding window size and a sliding step size, where the sliding window size is the number of data items contained in the window and the sliding step size is the number of data items moved in the window. Divide the data set into multiple subsets with the same data items according to the sliding window size and the sliding step size.

[0049] S6.3. Select the maximum value of each data subset, sum it up and take the average, and the obtained average value is the health threshold.

[0050] The technical concept of the present invention is: using the data collected by the wind turbine SCADA system, with the wind turbine generator, gearbox and converter as the monitoring objects, the MQE value is used as the performance indicator of the present invention to characterize the health status of the wind turbine. First, data selection is performed to select the sensor data with high correlation within the generator, gearbox and converter components. Secondly, data processing is performed to clean and normalize the selected data. The data characterizing the components is then input into the LSTM to obtain the predicted value and calculate the residual. The data residual of each component is then input into the underlying SOM model of the MLSOM to obtain the MQE of each component. i Value and health threshold; MQE of the three components i The values ​​are combined to form new data as the input data of the top-level SOM model, which is then input into the top-level SOM model in the MLSOM to obtain the normal behavior model and health threshold of the whole machine. The normal behavior model is used to predict subsequent data, and the operating status of the wind turbine is judged by the health threshold of the whole machine. If a fault alarm occurs in the wind turbine, the health threshold of different components is combined with the MQE of each component. i The wind turbine components with abnormal conditions can be determined by comparing the values.

[0051] The beneficial effects of the present invention are: using the LSTM model to model and predict the data under normal operating conditions of the unit, capturing the nonlinear relationship between historical wind speed, active power, temperature and other data, and subtracting the predicted value calculated by the LSTM model from the actual value to obtain the residual, and using the residual as a new input of the SOM model, so that the SOM model can better evaluate the operating status of the wind turbine; using MLSOM to fuse the information of different components to realize the whole machine status monitoring of the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is the overall framework diagram of a wind turbine condition monitoring method based on LSTM-MLSOM;

[0053] Figure 2 It is a three-layer MLSOM model structure diagram;

[0054] Figure 3 This is the structure diagram of the MLSOM model of the wind turbine;

[0055] Figure 4 The wind speed and power curves of wind turbines before and after data cleaning;

[0056] Figure 5 Constructing diagrams to model normal wind turbine behavior;

[0057] Figure 6 This is the result diagram of wind turbine component-level condition monitoring;

[0058] Figure 7 This is the result diagram of wind turbine complete machine status monitoring. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings.

[0060] Reference Figures 1 to 7 A wind turbine condition monitoring method based on LSTM-MLSOM is proposed. The overall framework of this method is shown in the attached Figure 1 As shown, the following steps are included:

[0061] S1. Obtain sensor data with high correlation with generator active power from different components of the wind turbine from the wind farm SCADA system (including sensor data with high correlation within the generator, gearbox, and converter components);

[0062] S2. Taking the generator components of a wind turbine as an example, a training data set and a test data set of the generator components are established, and the two data sets are preprocessed;

[0063] Data preprocessing includes data cleaning and normalization. First, in the data cleaning stage, missing data, data with wind turbine operating power of 0, data outside the normal operating range, data with limited power operation, and outlier data are deleted. Then, the data is normalized. The normalization formula is as follows:

[0064]

[0065] S3. Inputting the preprocessed training data set into the LSTM model to obtain a trained LSTM model;

[0066] The loss function L used in the LSTM model MSE for

[0067]

[0068] S4. Input the training data set into the trained LSTM model to obtain the predicted value, and then subtract the predicted value from the actual value at the same time to obtain the residual value;

[0069] S5. Perform S1-S4 on the wind turbine gearbox and converter to obtain the LSTM models and residual values ​​of the gearbox and converter;

[0070] S6. Since the SCADA system does not contain data representing individual components or the entire wind turbine, it is not applicable to the original MLSOM model (see Appendix 1). Figure 2 Therefore, a MLSOM model for wind turbines is proposed (as shown in the attached Figure 3As shown in the figure), first, the residual values ​​of different components are input into the bottom SOM in the MLSOM to build the normal behavior model of the components and obtain the trained SOM models corresponding to the different components;

[0071] The Euclidean distance between the input vector and each neuron vector is calculated using the following formula:

[0072]

[0073] The neuron update formula of the SOM model is as follows:

[0074] w j (t+1)=w j (t)+η(t)h i,j (t)(X i -w j (t)) (4)

[0075] S7. Calculate the MQE of different component training data through the trained SOM models corresponding to different components i value, and through the different components of MQE i The value determines the health threshold of the corresponding component;

[0076] The calculation formula of MQE value is as follows:

[0077] MQE=||X i -W BMU || (5)

[0078] S8, MQE of different components at the same time i The value combination generates new data, which is input into the top SOM in MLSOM to build the normal behavior model of the whole machine (such as the attached Figure 5 As shown), and calculate the MQE of the whole machine WT value;

[0079] S9, MQE based on the top-level SOM model WT The value is used to determine the health threshold of the whole machine, so as to judge whether the wind turbine is in normal operation;

[0080] S10, input the test data sets of different components of the wind turbine into the corresponding trained LSTM model, output the predicted value, calculate the residual value, and then input the residual value into the trained SOM model to obtain the MQE corresponding to different components i Value (if attached) Figure 6 As shown), MQE i The values ​​are combined and input into the trained top-level SOM model to obtain the MQE of the whole machine. WT Value (if attached) Figure 7 When the MQE of the wind turbineWT If the value exceeds the health threshold of the whole machine for n times continuously, it is determined that the operation state of the wind turbine is abnormal. Then, the MQE calculated by the test data sets of different components is used to calculate the MQE value. i The value is compared with the health threshold of the corresponding component. If the MQE of the component i If the value exceeds the corresponding health threshold value n times continuously, the abnormal component in the wind turbine can be determined and repaired and maintained.

[0081] Example:

[0082] This paper uses online SCADA data from a doubly-fed asynchronous wind turbine in southwest China to verify the wind turbine condition monitoring method based on LSTM-MLSOM. The specific process is as follows:

[0083] 1) Sensor data with a high correlation with the active power of the generator from different components of the wind turbine are obtained from the wind farm's SCADA system. The sensor data selected for the generator component include active power, winding temperature, non-drive end bearing temperature, and generator speed; the sensor data selected for the converter component include grid current, motor speed, and machine-side module temperature; and the sensor data selected for the gearbox include pump inlet pressure, oil temperature, front end bearing temperature, and pump outlet pressure.

[0084] 2) Taking the generator components of wind turbines as an example, a training data set and a test data set of the generator components are established, and abnormal data cleaning and normalization are performed on the two data sets (see the attached Figure 4 shown).

[0085] 3) Input the processed data into LSTM to obtain the predicted value, subtract it from the actual value, and obtain the residual value.

[0086] 4) Perform the same steps on the wind turbine gearbox and converter to obtain the residual value; input the residual value of each component into the bottom SOM in MLSOM to obtain the trained SOM model and MQE corresponding to different components i Value, through MQE of different components i The health threshold of the corresponding component is determined by the value; then the MQE of different components at the same time is calculated. i The value combination generates new data, which is input into the top SOM in MLSOM to build a normal behavior model (such as the attached Figure 5 As shown), obtain the trained top-level SOM model and MQE WT value, and through MQE WT The value determines the health threshold of the entire machine.

[0087] 5) Input the test data into the LSTM-MLSOM model to obtain the wind turbine components (such as the attached Figure 6as shown) and the status monitoring results of the whole machine (as shown in the attached Figure 7 As shown in the figure, it can be seen that wind turbines had already experienced a lot of MQE at the end of January 2017. WT The value exceeds the threshold. In the status monitoring results of the generator components, it can also be seen that at the end of January 2017, many MQE values ​​exceeded the threshold, indicating that the generator components were abnormal. Compared with the SCADA system, an early warning was given, which effectively verified the wind turbine status monitoring method based on LSTM-MLSOM.

[0088] The above embodiments are only preferred embodiments of the present invention and are not limitations on the technical solutions of the present invention. Any technical solution that can be implemented on the basis of the above embodiments without creative work should be deemed to fall within the scope of protection of the patent of the present invention.

Claims

1. A wind turbine state monitoring method based on LSTM-MLSOM, characterized in that: The following steps are involved: S1. Obtain sensor data of different components of the wind turbine related to the generator active power from the wind farm’s SCADA system; S2. Using historical data of the above-mentioned different components when the wind turbine is generating electricity normally within a predetermined wind speed range and without any faults as a training data set, wherein each component corresponds to a training data set, and using online operation data as a test data set, wherein each component corresponds to a test data set, and preprocessing the above two data sets; Normal power generation without faults means that the wind turbine operating data are within the set range; S3. Input the preprocessed training data sets of different components into the corresponding LSTM models respectively. Use the LSTM models to retain the nonlinear relationship between different data in each component and discard irrelevant noise information in the data, so as to obtain the trained LSTM models of different components. S4. Input the preprocessed training data sets of different components into the trained LSTM models of different components respectively to obtain the predicted values ​​of different components, and then subtract the predicted values ​​of different components from the actual values ​​of the corresponding components at the same time to obtain the residual values ​​of different components; S5. Input the residual values ​​of different components into the SOM model corresponding to the bottom layer in the MLSOM to construct normal behavior models of different components. The normal behavior model refers to a mathematical model that can correctly describe the relationship between various parameters and performance of the device when it meets the design requirements and achieves the expected functions. The trained SOM model corresponding to the different components is the normal behavior model. S6. Calculate the MQE of different component training data through the trained SOM models corresponding to different components i value, and through the different components of MQE i The value determines the health threshold of the corresponding component; S7, MQE of different components at the same time i The value combination generates new data, which is input into the top SOM in the MLSOM to build the normal behavior model of the whole machine and calculate the MQE of the whole machine. WT value; S8. MQE based on the top-level SOM model WT The value is used to determine the health threshold of the whole machine, so as to judge whether the wind turbine is in normal operation; S9. Input the test data sets of different components of the wind turbine into the corresponding trained LSTM model, output the predicted value, calculate the residual value, and then input the residual value into the trained SOM model to obtain the MQE of different components. i value, MQE i The values ​​are combined and input into the trained top-level SOM model to obtain the MQE of the wind turbine. WT Value; When the MQE of the wind turbine WT If the value exceeds the health threshold of the whole machine for n times continuously, it is determined that the operation status of the wind turbine is abnormal; Afterwards, the MQE obtained by calculating the test data sets of different components i The value is compared with the health threshold of the corresponding component. If the MQE of the component i If the value exceeds the corresponding health threshold value n times continuously, the faulty component in the wind turbine can be determined and repaired and maintained.

2. A wind turbine state monitoring method based on LSTM-MLSOM according to claim 1, characterized in that: In step S2, the process of preprocessing the acquired wind turbine data is as follows: S2.

1. Clear abnormal data in historical data; S2.

2. To eliminate the dimension effect, the cleaned data is normalized.

3. A wind turbine state monitoring method based on LSTM-MLSOM according to claim 2, characterized in that: In step S2, abnormal data includes missing value data, shutdown data, power-limited data, data exceeding the normal operating range of the wind turbine, and outlier data. To eliminate the dimensionality effect, the cleaned data is normalized. Normalization is to convert the original data range to the [0, 1] interval. The normalization formula is as follows: Where x i,j norm Represents the normalized value, x i max and x i min Represents the maximum and minimum values ​​of the i-th feature quantity, x i,j Represents the jth data point of the i-th feature value.

4. The wind turbine state monitoring method based on LSTM-MLSOM according to claim 1, characterized in that: The step S3 includes the following steps: S3.

1. Set the length of the dataset input sequence and the time step, where the length of the dataset input sequence is the time window size and the time step is the time interval between samples. Construct multiple data subsets of the same size based on the input sequence length and time step. S3.

2. Set the input feature dimension, number of hidden layers, number of hidden layer neurons, output layer dimension, training cycle, optimization algorithm, and loss function in the LSTM model. S3.

3. Select a data subset according to the time series and input it into the LSTM model to output the predicted value; S3.

4. Calculate a loss function value based on the predicted value, and update the model parameters according to the loss function value; S3.

5. Select the next data subset according to the time step and execute steps S3.3-S3.4 until the training cycle reaches the set number of times.

5. The wind turbine state monitoring method based on LSTM-MLSOM according to claim 4, characterized in that: The loss function L used in the LSTM model in S3.2 MSE for Where m represents the total number of data values, y i represents the true value, Represents the predicted value.

6. A wind turbine state monitoring method based on LSTM-MLSOM according to claim 1, characterized in that: The step S5 includes the following steps: S5.

1. Set the number of neurons, learning rate, neighborhood radius, and number of iterations in the competitive layer of the SOM network model. S5.

2. Randomly select a data and input it into the SOM network model; S5.

3. Calculate the Euclidean distance between each neuron and the input data, and find the neuron with the smallest distance to the input data. This neuron is called the winning neuron. S5.

4. A neighborhood is determined with the winning neuron as the center, based on the set neighborhood radius. All neurons in the neighborhood are updated according to the update formula. The completion of the update indicates that one training is complete. S5.

5. Repeat S5.1-S5.4 until the training times reach the preset times and the training ends.

7. A wind turbine state monitoring method based on LSTM-MLSOM according to claim 6, characterized in that: The Euclidean distance calculation formula in S5.3 is: Where d represents the Euclidean distance, x i Represents the i-th dimension value of x, y i Represents the i-th dimension value of y.

8. The wind turbine state monitoring method based on LSTM-MLSOM according to claim 6, characterized in that: The update formula w used by the SOM network model in S5.4 j (t+1) is: w j (t+1)=w j (t)+η(t)h i,j (t)(X i -w j (t)) (4) Where w j (t+1) represents the updated weight, w j (t) represents the unupdated weight, η(t) represents the learning rate, and h i,j (t) is the nearest neighbor function, X i is the input vector.

9. The wind turbine state monitoring method based on LSTM-MLSOM according to claim 1, characterized in that: In step S6, the calculation formula of the MQE value is as follows: MQE=||X i -W BMU || (5) Where, X i is the input vector, W BMU is the winning neuron weight vector, ||X i -W BMU || means calculating the Euclidean distance between the input vector and the winning neuron weight.

10. The wind turbine state monitoring method based on LSTM-MLSOM according to claim 1, characterized in that: In step S6 or S8, the health threshold is determined as follows: S6.

1. Construct a data set from the calculated MQE values; S6.

2. Set a sliding window size and a sliding step size, where the sliding window size is the number of data items contained in the window and the sliding step size is the number of data items moved in the window. Divide the data set into multiple subsets with the same data items according to the sliding window size and the sliding step size. S6.

3. Select the maximum value of each data subset, sum it up and take the average, and the obtained average value is the health threshold.

Images