A wind turbine fault early warning method based on operation data

Abstract

The present invention proposes a wind turbine fault warning method based on operation data, including: S1, obtaining historical data of wind turbines; S2, processing daily average, maximum and minimum values; S3, dividing the data set; S4, and constructing a new training set; S5, constructing a long short-term memory network LSTM model and training to obtain the best prediction model; S6, inputting the characteristic data of the day to be predicted to obtain the final prediction result. The fault warning system can detect potential abnormal trends by comparing historical data and real-time data under normal operating conditions, and warn of possible faults of the unit in advance. The warning information can be sent to the operation and maintenance personnel through the remote monitoring system to achieve early intervention of the fault and avoid equipment damage and shutdown. The present invention significantly reduces the unplanned shutdown rate of wind turbines, extends the service life of equipment, and optimizes the operation and maintenance costs of wind farms by effectively predicting and identifying potential faults.

Description

Technical Field

[0001] The present invention relates to the field of wind power technology, and in particular to a wind turbine fault early warning method based on operation data. Background Art

[0002] In the wind power industry, the operational reliability and maintenance efficiency of wind turbines are important indicators for measuring the operational level of wind farms. With the rapid development of wind power technology and the continuous increase in wind power installed capacity, the scale of wind turbines is increasing and the structure is becoming increasingly complex, which puts higher requirements on the operation and maintenance management of wind turbines. Traditional operation and maintenance methods often rely on regular inspections and post-failure repairs. This method is not only inefficient, but also difficult to effectively prevent the occurrence of failures, resulting in a high rate of unplanned downtime of wind turbines, shortened equipment life, and high operation and maintenance costs.

[0003] In order to improve the operational reliability and maintenance efficiency of wind turbines, the existing technology has begun to try to use the operating data of wind turbines for fault warning. These systems usually capture the operating status information of the unit by real-time monitoring of key parameters during the operation of wind turbines, such as wind speed, wind direction, vibration, speed, temperature, power generation, etc. However, these systems still have some shortcomings in practical applications, including: 1. The accuracy of fault warning needs to be improved: Most of the existing fault warning systems are based on simple threshold judgment or statistical methods, which are difficult to accurately capture the complex fault characteristics during the operation of the unit, resulting in low warning accuracy, false alarms and missed alarms. 2. The fault prediction model lacks adaptability: Due to the complexity and diversity of the operating environment of wind turbines, there are differences in the fault characteristics between different units and different components. The existing fault prediction models often lack sufficient adaptability and are difficult to make accurate predictions for different units and components. 3. Limited data analysis and processing capabilities: With the continuous accumulation of wind turbine operation data, the amount of data has shown explosive growth. The existing data analysis and processing algorithms have performance bottlenecks when processing large-scale data, and it is difficult to strike a balance between real-time, accuracy and robustness. 4. Imperfect early warning information transmission and response mechanism: Most existing fault early warning systems use local monitoring methods, and early warning information is difficult to transmit to remote operation and maintenance personnel in a timely manner. At the same time, the response mechanism for early warning information is not perfect, and there is a lack of effective fault handling suggestions and emergency response measures.

[0004] Therefore, the operation, maintenance and early warning of wind turbines should continue to be monitored. Summary of the invention

[0005] The present invention provides a wind turbine fault warning method based on operating data, which detects potential abnormal trends by comparing historical data and real-time data under normal operating conditions, and warns of possible faults of the wind turbine in advance. To solve the above technical problems, the technical solution adopted by the present invention is: A wind turbine fault early warning method based on operation data, comprising: S1. Obtain historical data of wind turbine generator sets; the historical data includes time, active power, wind speed, wind direction, turbulence intensity, temperature, and fault information and type of the unit controller and perform preprocessing; S2. Process the daily active power, wind speed, wind direction, turbulence intensity and temperature in the data into daily average, maximum and minimum values ​​to represent the characteristic type of the operating parameters of the day; S3, standardizing the characteristic type data of the operating parameters, and using the Kmeans algorithm to perform clustering on the standardized data and divide the data set; S4. Select the corresponding data set according to the characteristic type of the operating parameters of the day to be predicted, search for similar unit controller fault information on the predicted day and historical data based on weighted grey correlation analysis, and construct a new training set; S5. Build a long short-term memory network LSTM model, and use active power, wind speed, wind direction, turbulence intensity, temperature and similar unit controller fault information as input training, use wind turbine fault information as output training LSTM model, and use grid search method to adjust parameters to obtain the best prediction model; S6. Input the characteristic data of the day to be predicted to obtain the final prediction result.

[0006] The basis for dividing the operating parameter feature type data set in the above S3 is to select data of different degrees of wind speed, turbulence intensity and temperature according to time information and local historical data information and combine them into different data combinations.

[0007] The specific method of S4 mentioned above includes: S4.1. Construct a sample sequence Xi(k1, k2, ...kn) for each moment of the data set, where n is the window length of the data sequence, and k1, k2, ...kn represent the active power, wind speed, wind direction, turbulence intensity and temperature data at moment i, respectively, and normalize the data to form a comparison series; S4.2. Calculate the weighted correlation degree for each feature, select the feature data at a certain moment as the reference sequence, perform grey correlation analysis with all moments in the data set, and calculate the sequence correlation coefficient between the comparison sequence and the reference sequence as follows: ; Xi(k) is the reference sequence, y(k) is the comparison sequence, and ξi represents the correlation coefficient of element k; minimink|y(k)-xi(k)| is the minimum of the absolute differences between all comparison sequences and the reference sequence; similarly, maximaxk|y(k)-xi(k)| is the maximum of the absolute differences between sequences; S4.3. Since the influence degrees of different weather characteristics on the output of wind turbines are different, weight calculation is introduced when calculating the grey correlation degree. The specific steps are as follows: S4.3.1. Construct a data matrix using the original dimensionless data ; where m represents the number of features of each group of sample data, and n represents the sample size of the data; S4.3.2. Determine the ideal value and negative ideal value of the sample; ; where ; S4.3.3. Use the Euclidean formula to calculate the distance from each element in a column of samples to the ideal value; , ; where represents the Euclidean distance from the sample data to the ideal value, and is the Euclidean distance from the data to the negative ideal value; S4.3.4. Comprehensive distance calculation; ; Calculate the closeness degree of each scheme to the ideal solution. When is larger, it means that the element is more important in the sample data; S4.3.5. Vector weight calculation; ; is the weight vector of the sequence to be obtained; S4.4. Grey correlation degree analysis; after calculating the correlation coefficient for each element in Xi(k), the grey correlation degree r can be calculated through the following formula i : ; When ri>0.7, it is identified as the first-level correlation; when 0.5<ri<0.7, it is identified as the second-level correlation; when ri<0.5, it is the third-level correlation, where α(k) is the weight corresponding to the element; S4.5. Take r iThe values ​​are sorted from large to small, and the three largest values ​​in the first-level correlation value range are taken to extract the corresponding power data and construct a new data set.

[0008] In the above S5, parameters are set for data sets under different operating parameter feature types, including a short input time, the number of training area data samples, and the number of training times.

[0009] The above S5 introduces the convolutional neural network CNN for data feature extraction: ; In the formula, represents the mean of the data, represents the sample standard deviation; In the one-dimensional convolution process, the convolution kernel slides on each sample and performs convolution. The output dimension of the convolution layer is determined by the number of convolution kernels. Each sample is converted into a feature map. The formula is as follows: ; In the formula, represents the transpose of the convolution kernel matrix, represents a nonlinear activation function, represents the bias variable, Represents the sample sequence from arrive Finally, multiple feature maps are combined into a one-dimensional vector as the output of the convolutional neural network CNN.

[0010] The forget gate and input gate of the long short-term memory network LSTM in the above S5 are combined together, and the formula is as follows: ; in, is the output of the forget gate, is the current state, The value ranges from 0 to 1, which determines which values ​​are updated to generate the current state. is a new state candidate vector. The function of the output gate is to select part of the current state as the output: ; The output gate integrates the previous state and the current state as the current weight ; After a Activation function and Multiply to get the output signal; Finally, a fully connected layer is set after LSTM to combine the previously extracted feature data to generate output data. The formula is as follows: ; In the formula, is a nonlinear activation function, is the weight matrix of the fully connected layer, is the bias vector, is the input vector, which is the output vector of the LSTM network.

[0011] The present invention provides a wind turbine fault warning method based on operation data, which establishes a wind turbine fault prediction model by real-time monitoring and data collection of key parameters in the operation of the wind turbine, such as wind speed, wind direction, vibration, rotation speed, temperature, power generation, etc., and using advanced data analysis and machine learning algorithms. The fault warning system can detect potential abnormal trends by comparing historical data and real-time data under normal operating conditions, and warn of possible faults of the unit in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is the inventive flow chart of the present method. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following content will systematically and completely describe the specific technical solutions of the present invention in combination with the drawings provided according to the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0014] Embodiment 1: like Figure 1 As shown in , a wind turbine fault early warning method based on operation data includes: S1. Obtain historical data of wind turbine generator sets; the historical data includes time, active power, wind speed, wind direction, turbulence intensity, temperature, and fault information and type of the unit controller and perform preprocessing; S2. Process the daily active power, wind speed, wind direction, turbulence intensity and temperature in the data into daily average, maximum and minimum values ​​to represent the characteristic type of the operating parameters of the day; S3, standardizing the characteristic type data of the operating parameters, and using the Kmeans algorithm to perform clustering on the standardized data and divide the data set; S4. Select a corresponding data set according to the characteristic type of the operating parameters of the day to be predicted, search for similar unit controller fault information on the predicted day and historical data based on weighted grey correlation analysis, and construct a new training set; S5. Build a long short-term memory network LSTM model, and use active power, wind speed, wind direction, turbulence intensity, temperature and similar unit controller fault information as input training, use wind turbine fault information as output training LSTM model, and use grid search method to adjust parameters to obtain the best prediction model; S6. Input the characteristic data of the day to be predicted to obtain the final prediction result.

[0015] The basis for dividing the operating parameter feature type data set in the above S3 is to select data of different degrees of wind speed, turbulence intensity and temperature according to time information and local historical data information and combine them into different data combinations.

[0016] The specific method of S4 mentioned above includes: S4.1. Construct a sample sequence Xi(k1, k2, ...kn) for each moment of the data set, where n is the window length of the data sequence, and k1, k2, ...kn represent the active power, wind speed, wind direction, turbulence intensity and temperature data at moment i, respectively, and normalize the data to form a comparison series; S4.2. Calculate the weighted correlation degree for each feature, select the feature data at a certain moment as the reference sequence, perform grey correlation analysis with all moments in the data set, and calculate the sequence correlation coefficient between the comparison sequence and the reference sequence as follows:

[0017] Xi(k) is the reference sequence, y(k) is the comparison sequence, ξi represents the correlation coefficient of element k; minimink|y(k)-xi(k)| is the minimum absolute difference between all comparison sequences and the reference sequence; similarly maximaxk|y(k)-xi(k)| is the maximum absolute difference between the sequences; S4.3. Since different weather characteristics have different effects on wind turbine output, weight calculation is introduced when calculating grey correlation. The specific steps are as follows: S4.3.1. Constructing a data matrix using raw dimensionless data ; Where m represents the number of features of each set of sample data, and n represents the sample size of the data; S4.3.2, determine the sample ideal value and negative ideal value; ; in ; S4.3.3, use the Euclidean formula to calculate the distance between each element in a column of samples and the ideal value; , ; wherein represents the Euclidean distance from the sample data to the ideal value, is the Euclidean distance from the data to the negative ideal value; S4.3.4, Comprehensive distance calculation; ; Calculate the closeness degree of each scheme to the ideal solution. When is larger, it means that the element is more important in the sample data; S4.3.5, Vector weight calculation; ; is the weight vector of the sequence to be obtained; S4.4, Grey relational analysis; After calculating the correlation coefficient for each element in Xi(k), the grey relational degree r can be calculated through the following formula i : ; When ri > 0.7, it is identified as the first-level correlation; when 0.5 < ri < 0.7, it is identified as the second-level correlation; when ri < 0.5, it is the third-level correlation, where α(k) is the weight corresponding to the element; S4.5, Sort the r i values from large to small, take the three largest values within the range of the first-level correlation degree, extract the corresponding power data, and construct a new data set.

[0018] In the above S5, parameter settings are respectively performed on the data sets under different operating parameter characteristic types, including the input time length, the number of data samples in the training area, and the number of training times.

[0019] In the above S5, a convolutional neural network CNN is introduced for data feature extraction: ; In the formula, represents the mean value of the data, represents the sample standard deviation; In the one-dimensional convolution process, the convolution kernel slides on each sample and performs convolution. The output dimension of the convolution layer is determined by the number of convolution kernels, and each sample is transformed into a feature map. The formula is as follows: ; In the formula, represents the transpose of the convolution kernel matrix, represents the non-linear activation function, represents the bias variable, represents the sample sequence from arrive Finally, multiple feature maps are combined into a one-dimensional vector as the output of the convolutional neural network CNN.

[0020] The forget gate and input gate of the long short-term memory network LSTM in the above S5 are combined together, and the formula is as follows: ; in, is the output of the forget gate, is the current state, The value ranges from 0 to 1, which determines which values ​​are updated to generate the current state. is a new state candidate vector. The function of the output gate is to select part of the current state as the output: ; The output gate integrates the previous state and the current state as the current weight ; After a Activation function and Multiply to get the output signal; Finally, a fully connected layer is set after LSTM to combine the previously extracted feature data to generate output data. The formula is as follows: ; In the formula, is a nonlinear activation function, is the weight matrix of the fully connected layer, is the bias vector, is the input vector, which is the output vector of the LSTM network.

[0021] When constructing the LSTM network model, the Dropout and Dense values ​​are set to the commonly used values ​​of 0.2 and 1, and the active power, wind speed, wind direction, turbulence intensity, temperature, and similar unit controller fault information are used as input training, and the fault information of the wind turbine is used as output training of the LSTM model. Under different operating parameter characteristics, the output of wind turbines is different. Therefore, setting parameters for different operating parameter characteristics can enable LSTM to better learn data patterns and improve prediction performance. The input time step can determine how many time nodes are used to predict the results at the next time point. This model adjusts the input time step of input_shape, with an interval of 5, and tests the optimal parameters in the interval [5,10,15,…,30]. The grid search method is used to adjust the batch_size and epochs parameters of LSTM. The batch_size and epochs are both intervals of 20, and the optimal parameter search is performed in the interval [10,30,50,…,100]. For the cloudy data set, the optimal parameter of input_shape time step is 5, batch_size and epochs are set to 10 and 50 respectively, which is the best prediction model obtained, and the final prediction result is obtained with this parameter.

Claims

1. A wind turbine fault early warning method based on operation data, characterized in that: include: S1. Obtain historical data of wind turbine generator sets; the historical data includes time, active power, wind speed, wind direction, turbulence intensity, temperature, and fault information and type of the unit controller and perform preprocessing; S2. Process the daily active power, wind speed, wind direction, turbulence intensity and temperature in the data into daily average, maximum and minimum values ​​to represent the characteristic type of the operating parameters of the day; S3, standardizing the characteristic type data of the operating parameters, and using the Kmeans algorithm to perform clustering on the standardized data and divide the data set; S4. Select the corresponding data set according to the characteristic type of the operating parameters of the day to be predicted, search for similar unit controller fault information on the predicted day and historical data based on weighted grey correlation analysis, and construct a new training set; S5. Build a long short-term memory network LSTM model, and use active power, wind speed, wind direction, turbulence intensity, temperature and similar unit controller fault information as input training, use wind turbine fault information as output training LSTM model, and use grid search method to adjust parameters to obtain the best prediction model; S6. Input the characteristic data of the day to be predicted to obtain the final prediction result.

2. According to the wind turbine fault early warning method based on operation data as described in claim 1, it is characterized in that: The basis for dividing the operating parameter characteristic type data set in S3 is to select data of different degrees of wind speed, turbulence intensity and temperature according to time information and local historical data information and combine them into different data combinations.

3. According to the wind turbine fault early warning method based on operation data as described in claim 2, it is characterized in that: The specific method of S4 includes: S4.

1. Construct a sample sequence Xi(k1, k2, ...kn) for each moment of the data set, where n is the window length of the data sequence, and k1, k2, ...kn represent the active power, wind speed, wind direction, turbulence intensity and temperature data at moment i, respectively, and normalize the data to form a comparison series; S4.

2. Calculate the weighted correlation degree for each feature, select the feature data at a certain moment as the reference sequence, perform grey correlation analysis with all moments in the data set, and calculate the sequence correlation coefficient between the comparison sequence and the reference sequence as follows: ； Xi(k) is the reference sequence, y(k) is the comparison sequence, ξi represents the correlation coefficient of element k; minimink|y(k)-xi(k)| is the minimum absolute difference between all comparison sequences and the reference sequence; similarly maximaxk|y(k)-xi(k)| is the maximum absolute difference between the sequences; S4.

3. Since different weather characteristics have different effects on wind turbine output, weight calculation is introduced when calculating grey correlation. The specific steps are as follows: S4.3.

1. Constructing a data matrix using raw dimensionless data ； Where m represents the number of features of each set of sample data, and n represents the sample size of the data; S4.3.2, determine the sample ideal value and negative ideal value; ； in ; S4.3.3, use the Euclidean formula to calculate the distance between each element in a column of samples and the ideal value; ， ； in Represents the Euclidean distance from the sample data to the ideal value, is the Euclidean distance from the data to the negative ideal value; S4.3.4, comprehensive distance calculation; ； Calculate the closeness of each solution to the ideal solution. The larger it is, the more important the element is in the sample data; S4.3.5, vector weight calculation; is the weight vector of the required sequence; S4.4, Grey correlation analysis: After calculating the correlation coefficient of each element in Xi(k), the grey correlation r can be calculated by the following formula: i : ； If ri > 0.7, it is recognized as the first - level relevance; if 0.5 < ri < 0.7, it is recognized as the second - level relevance; if ri < 0.5, it is the third - level relevance, where α(k) is the weight corresponding to the element. S4.

5. i The values ​​are sorted from large to small, and the three largest values ​​in the first-level correlation value range are taken to extract the corresponding power data and construct a new data set.

4. A wind turbine fault early warning method based on operation data according to claim 3, characterized in that: In S5, parameter settings are separately performed on data sets under different operating - parameter feature types, including input time, the number of data samples in the training area, and the number of training times.

5. A wind turbine fault early warning method based on operation data according to claim 4, characterized in that: In S5, a convolutional neural network CNN is introduced for data - feature extraction: ； In the formula, represents the mean of the data, represents the sample standard deviation; In the one - dimensional convolution process, the convolutional kernel slides on each sample and performs convolution. The output dimension of the convolutional layer is determined by the number of convolutional kernels, and each sample is transformed into a feature map. The formula is as follows: ； In the formula, represents the transpose of the convolution kernel matrix, represents a nonlinear activation function, represents the bias variable, Represents the sample sequence from arrive Finally, multiple feature maps are combined into a one-dimensional vector as the output of the convolutional neural network CNN.

6. A wind turbine fault early warning method based on operation data according to claim 5, characterized in that: In S5, the forget gate and input gate of the long short - term memory network LSTM are combined together. The formula is as follows: ； in, is the output of the forget gate, is the current state, The value ranges from 0 to 1, which determines which values ​​are updated to generate the current state. is a new state candidate vector. The function of the output gate is to select part of the current state as the output: ； The output gate integrates the previous state and the current state as the current weight ; After a Activation function and Multiply to get the output signal; Finally, a fully - connected layer is set after the LSTM to combine the previously extracted feature data to generate output data. The formula is as follows: ； In the formula, is a nonlinear activation function, is the weight matrix of the fully connected layer, is the bias vector, is the input vector, which is the output vector of the LSTM network.

Images