A Wind Turbine Condition Monitoring Method Based on the GAN-QP Feature Migration Model
Through the combination of GAN-QP feature migration model and autoencoder, the status monitoring problem of newly installed or data loss wind turbines is solved, and effective monitoring and efficient operation and maintenance are achieved in the case of insufficient data.
Patent Information
- Application Number
- CN202211739162.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-31
- Publication Date
- 2025-07-18
- Estimated Expiration
- 2042-12-31
AI Technical Summary
The state monitoring model cannot be effectively trained in the wind turbines with newly installed or data loss, mainly because the training data does not meet the requirements of independent and same distribution.
The GAN-QP feature migration model is used to migrate the data feature distribution of the target domain unit to the source domain unit, and combine the autoencoder training model to realize data distribution adaptation, and the unit status is judged by means of mean square error.
Under the condition of insufficient data, the wind turbine status is effectively monitored, which reduces the early warning frequency and improves operation and maintenance efficiency.
Smart Images
Figure CN116011332B_ABST
Abstract
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 a GAN-QP feature migration model. Background Art
[0002] With the development goals of "carbon peak and carbon neutrality", clean and renewable energy represented by wind energy and solar energy have ushered in new development opportunities. In particular, wind power generation technology has been continuously developed and improved, and the newly installed capacity of wind turbines has steadily increased. Accordingly, in order to ensure the safe and reliable operation of wind turbines in complex and changing environments, the research on intelligent condition monitoring technology for wind turbines has become a hot issue in the wind power industry. Thanks to the development of artificial intelligence technology, the intelligent condition monitoring method of wind turbines based on data acquisition and supervisory control (SCADA) data has gradually replaced traditional condition monitoring methods such as vibration monitoring and oil monitoring and become the mainstream method. This method usually combines normal behavior modeling with residual analysis to monitor the condition of wind turbines, but the training of the model requires sufficient training data, and the training data and test data must obey the independent and identically distributed conditions. Therefore, it is impossible to establish an effective model for newly installed units or units with lost data records due to insufficient training data. This method has certain limitations.
[0003] Transfer learning provides a new idea for establishing a wind turbine condition monitoring model under the condition of insufficient data. It can extract transferable features or knowledge structures from source domains with sufficient relevant data and rich domain knowledge to improve the performance of the model in different but related target domains. To a certain extent, it relaxes the requirement that training data must be sufficient and must be independent and identically distributed with test data. Summary of the invention
[0004] The purpose of the present invention is to provide a wind turbine state monitoring method based on a feature migration model. The feature-based migration method can still realize knowledge reuse when the data distribution difference between the source domain and the target domain is large. Further combining it with adversarial learning can effectively promote the distribution adaptation between the source domain and the target domain data. To a certain extent, it relaxes the requirement that the training data must be sufficient and must obey the independent and identical distribution with the test data. It provides a new idea for establishing a wind turbine state monitoring model under insufficient data conditions for newly installed units or units with lost data records.
[0005] To achieve the above objectives, the present invention adopts the following technical solutions:
[0006] A wind turbine condition monitoring method based on a GAN-QP feature migration model comprises the following steps:
[0007] S1. Collect the historical data sets of the operating states of the generator systems of two wind turbines with the same type but different geographical locations and rated powers.
[0008] S2. Take one of the wind turbines as the source domain unit and the other as the target domain unit, and preprocess the historical data sets of the operating states of the generator systems of the two wind turbines respectively to obtain a source domain training set and a target domain training set.
[0009] S3. Train a feature transfer model based on the target domain training set and the source domain training set to obtain a trained feature transfer model, which can transform the feature distribution of the operating data of the target domain unit to the feature distribution of the operating data of the source domain unit.
[0010] S4. Train an autoencoder based on the source domain training set to learn the feature distribution rules of the data under the normal operating state of the source domain unit to obtain a trained autoencoder.
[0011] S5. Preprocess the online data set of the operating state of the generator system of the target domain unit to obtain a target domain test set, and transform the target domain test set to the source domain space based on the trained feature transfer model to obtain a source domain test set.
[0012] S6. Use the source domain test set obtained in step S5 as the input data of the trained autoencoder, calculate the mean square error between the source domain test set and the source domain training set, and judge whether the operating state of the generator system is normal based on the mean square error.
[0013] As a preferred solution, in step S2, the preprocessing of the historical data sets of the operating states of the generator systems of the source domain unit and the target domain unit includes the following steps:
[0014] S2.1. Delete the abnormal data in the historical data set.
[0015] S2.2. Normalize the historical data set processed in step S2.1.
[0016] As a preferred solution, in step S2.1, the abnormal data includes data with missing values, shutdown data, data outside the normal operating wind speed range of the wind turbine, and outlier data.
[0017] As a preferred solution, the judgment of the outlier data includes the following steps:
[0018] A. Select a certain data point in the historical data set.
[0019] B. Select multiple target data points in the historical data set that are closest to the data point selected in step A.
[0020] C. Calculate the average chain distance of the selected data point based on the selected data point in step A and multiple target data points;
[0021] D. Calculate the distance based on the connection anomaly factor of the selected data point based on the average chain distance of the selected data point and the average chain distances corresponding to the multiple target data points respectively;
[0022] E. Determine whether the selected data point is an outlier based on the distance based on the connection anomaly factor of the selected data point.
[0023] As a preferred solution, in step D, the following steps are included:
[0024] Calculate the average chain distances corresponding to the multiple target data points respectively;
[0025] Calculate the average value of the average chain distances of the multiple target data points;
[0026] Calculate the distance based on the connection anomaly factor of the selected data point based on the average chain distance of the selected data point and the average value of the average chain distances of the multiple target data points.
[0027] As a preferred solution, the average chain distance of the selected data point The calculation formula is:
[0028]
[0029] where k is the total number of target data points, and e i represents the distance from the i-th target data point to the selected data point.
[0030] As a preferred solution, the distance based on the connection anomaly factor COF k (p) of the selected data point has the following calculation formula:
[0031]
[0032] where N k (p) represents a set of multiple target data points, represents the average chain distance of the o-th target data point.
[0033] As a preferred solution, in step S3, the following steps are included:
[0034] S3.1. Divide the target domain training set into multiple batches;
[0035] S3.2. Select a batch of the target domain training set and input it into the generator network for preprocessing to obtain a generated data set;
[0036] S3.3. Input the generated dataset and the source domain training set into the discriminator network to output the discrimination results of the two;
[0037] S3.4. Calculate the loss function values of the generator network and the discriminator network based on the discrimination results, and update the model parameters of the generator network and the discriminator network based on the loss function values;
[0038] S3.5. Loop through steps S3.2 to S3.4 for a preset number of times, so that the feature distribution of the generated dataset of the generator network approximates the feature distribution of the source domain training set, and use the finally obtained generator network as the trained feature transfer model.
[0039] As a preferred solution, the loss function L used by the discriminator network D is:
[0040]
[0041] where D(·) represents the discriminator output, x g represents the data of the generated dataset, x r represents the data of the source domain training set, λ is a balance coefficient, d(x r , x g ) represents the Euclidean distance between x r and x g , p r is the data distribution of the source domain training set, p g is the data distribution of the generated dataset, is to take the expected value of each sample data in the data distribution of the source domain training set and the data distribution of the generated dataset.
[0042] As a preferred solution, the loss function L used by the generator network G is:
[0043]
[0044] The beneficial effects of the present invention are:
[0045] Aiming at the problem that the operation data of newly installed or data-lost wind turbines is insufficient and an effective traditional machine learning monitoring model cannot be trained, the present invention proposes a wind turbine condition monitoring method based on the GAN-QP feature transfer model. By using the trained feature transfer model, the feature distribution of the operation data of the target domain unit is transformed into the feature distribution of the operation data of the source domain unit. Furthermore, a self-encoder model with excellent performance can be trained using a large amount of data from the source domain unit, and the online dataset of the target domain unit transformed into the source domain space based on the trained feature transfer model is monitored to determine whether the generator system is operating normally.
[0046] The present invention provides a new idea for establishing an effective state monitoring model for mechanical equipment under the condition of data loss. Description of the Drawings
[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.
[0048] Figure 1 It is a flowchart of a wind turbine state monitoring method based on a GAN-QP feature transfer model.
[0049] Figure 2 It is a schematic diagram of the distribution adaptation of the source domain and target domain unit data based on the feature transfer model.
[0050] Figure 3 They are scatter plots of unit wind speed - active power before and after preprocessing of the historical data of the generator systems of the source domain unit and the target domain unit.
[0051] Figure 4 They are scatter plots of wind speed - generator speed - active power before and after distribution adaptation of the historical data of the generator systems of the source domain unit and the target domain unit.
[0052] Figure 5 They are monitoring diagrams of the generator states of the target domain units under the condition of no migration and under the condition of migration using the method described in the present invention. Detailed Embodiments
[0053] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Each detail in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.
[0054] Referring to Figure 1 , a wind turbine state monitoring method based on a GAN-QP feature transfer model includes the following steps:
[0055] S1. Collect the historical data sets of the operating states of the generator systems of two wind turbines with the same unit type but different geographical locations and rated powers;
[0056] S2. Take one wind turbine as the source domain unit and the other wind turbine as the target domain unit, and preprocess the historical data sets of the operating states of the generator systems of the two wind turbines respectively to obtain a source domain training set and a target domain training set;
[0057] S3. Train a feature transfer model based on the target domain training set and the source domain training set to obtain a trained feature transfer model, which can transform the feature distribution of the operating data of the target domain unit to the feature distribution of the operating data of the source domain unit;
[0058] S4. Train an autoencoder based on the source domain training set to learn the feature distribution rules of the data under the normal operating state of the source domain unit to obtain a trained autoencoder;
[0059] S5. Preprocess the online data set of the operating state of the generator system of the target domain unit to obtain a target domain test set, and transform the target domain test set to the source domain space based on the trained feature transfer model to obtain a source domain test set;
[0060] S6. Use the source domain test set obtained in step S5 as the input data of the trained autoencoder, calculate the mean square error between the source domain test set and the source domain training set, and judge whether the operating state of the generator system is normal based on the mean square error.
[0061] Further, in step S1, the data related to the operating state of the generator system is extracted as shown in the following table:
[0062]
[0063]
[0064] Further, in step S2, the preprocessing of the historical data sets of the operating states of the generator systems of the source domain unit and the target domain unit includes the following steps:
[0065] S2.1. Delete the abnormal data in the historical data set;
[0066] S2.2. Normalize the historical data set processed in step S2.1 to eliminate the influence of the dimension, reduce the difficulty of model training, and improve the accuracy of the model.
[0067] Further, in step S2.1, the abnormal data includes data with missing values, shutdown data, data outside the normal operating wind speed range of the wind turbine, and outlier data.
[0068] Specifically:
[0069] The shutdown data are data with active power of 0; the data outside the normal operating wind speed range of the wind turbine are data with wind speed less than the cut-in wind speed (3 m / s) and data with wind speed greater than the cut-out wind speed (25 m / s); the power-limited operation data are data with a constant power distributed below the rated power and not varying with wind speed, mainly obtained by manual control and unable to correctly reflect the operation law of the unit.
[0070] Figure 3 (a) is the scatter plot of wind speed - active power of the source domain unit generator system state historical data before preprocessing; Figure 3 (b) is the scatter plot of wind speed - active power of the source domain unit generator system state historical data after preprocessing; Figure 3 (c) is the scatter plot of wind speed - active power of the target domain unit generator system state historical data before preprocessing; Figure 3 (d) is the scatter plot of wind speed - active power of the target domain unit generator system state historical data after preprocessing.
[0071] After preprocessing the data set, both the source domain training set and the target domain training set obtained are normal data sets.
[0072] Furthermore, the judgment of the outlier data includes the following steps:
[0073] A. Select a certain data point in the historical data set;
[0074] B. Select multiple target data points in the historical data set that are closest to the data point selected in step A;
[0075] C. Calculate the average chain distance of the selected data point based on the selected data point in step A and the multiple target data points;
[0076] D. Calculate the distance of the selected data point based on the connection anomaly factor based on the average chain distance of the selected data point and the average chain distances corresponding to the multiple target data points respectively;
[0077] E. Based on the distance of the selected data point based on the connection anomaly factor, judge whether the selected data point is outlier data.
[0078] Furthermore, in step D, it includes the following steps:
[0079] Calculate the average chain distances corresponding to the multiple target data points respectively;
[0080] Calculate the average value of the average chain distances of the multiple target data points;
[0081] Calculate the distance of the selected data point based on the connection anomaly factor based on the average chain distance of the selected data point and the average value of the average chain distances of the multiple target data points.
[0082] Further, the average chain distance of the selected data points The calculation formula is as follows:
[0083]
[0084] where k is the total number of target data points, and e i represents the distance from the i-th target data point to the selected data point.
[0085] Further, the distance COF(p) of the selected data points based on the connection anomaly factor k (p) The calculation formula is as follows:
[0086]
[0087] where N k (p) represents a set of multiple target data points, represents the average chain distance of the o-th target data point.
[0088] Specifically:
[0089] The value of COF is the possibility that point p is an outlier. When the value of COF is less than 1, it indicates that the point is a normal value. When the value of COF is greater than or equal to 1, it indicates that the point is an outlier.
[0090] Further, the formula for normalizing the historical data set processed in step S2.1 is as follows:
[0091]
[0092] where x is the data of the historical data set processed in step S2.1, min(x) is the minimum value of x, and max(x) is the maximum value of x.
[0093] Further, after obtaining the source domain and target domain training sets through data preprocessing, step S3 is executed. In step S3, the following steps are included:
[0094] S3.1. Divide the target domain training set into multiple batches;
[0095] S3.2. Select a batch of the target domain training set and input it into the generator network for preprocessing to obtain a generated data set;
[0096] S3.3. Input the generated data set and the source domain training set into the discriminator network to output the discrimination results of the two;
[0097] S3.4. Calculate the loss function values of the generator network and the discriminator network based on the discrimination results, and update the model parameters of the generator network and the discriminator network based on the loss function values;
[0098] S3.5. Execute steps S3.2 to S3.4 in a loop for a preset number of times, so that the feature distribution of the generated dataset by the generator network approximates the feature distribution of the source domain training set, and use the finally obtained generator network as the trained feature transfer model.
[0099] Furthermore, the loss function L adopted by the discriminator network D is:
[0100]
[0101] where D(·) represents the output of the discriminator, x g represents the data of the generated dataset, x r represents the data of the source domain training set, λ is a balance coefficient, d(x r , x g ) represents the Euclidean distance between x r and x g , p r is the data distribution of the source domain training set, p g is the data distribution of the generated dataset, is to take the expected value of each sample data in the data distribution of the source domain training set and the generated dataset.
[0102] Furthermore, the loss function L adopted by the generator network G is:
[0103]
[0104] Specifically:
[0105] Based on the trained feature transfer model, perform distribution adaptation on the data of the source domain and target domain units. The data distributions of the wind speed - generator speed - active power of the source domain unit and the target domain unit before distribution adaptation are shown in Figure 4 (a), and the data distributions of the wind speed - generator speed - active power of the source domain unit and the target domain unit after distribution adaptation are shown in Figure 4 (b).
[0106] Through the method of the present invention, the feature transfer model can be trained based on the target domain training set and the source domain training set. The obtained trained feature transfer model can transform the feature distribution of the target domain unit operation data to the feature distribution of the source domain unit operation data, which can effectively promote the distribution adaptation between the source domain and the target domain data, and relaxes to a certain extent the requirement that the training data must be sufficient and follow the independent and identical distribution with the test data, providing a new idea for establishing a wind turbine condition monitoring model for newly installed units or units with lost data records under insufficient data conditions.
[0107] More specifically:
[0108] The mean square error calculated in step S6 is used to determine whether the operation state of the generator system is normal.
[0109] Since the wind farm is generally located in remote areas such as mountains and coasts and is inconvenient to maintain, in order to avoid frequent alarms, a relatively loose early warning strategy is selected. The source domain test set obtained in step S5 is used as the input data of the trained autoencoder, the mean square error between the source domain test set and the source domain training set is calculated, and the maximum value of the obtained mean square error is used as the health threshold to reduce the early warning frequency and improve the operation and maintenance efficiency of the wind farm.
[0110] Since both the source domain training set and the target domain training set used are the health data of the unit, the mean square errors obtained with the health data as the input data are all determined to be normal values. Therefore, using the maximum value as the monitoring threshold can reduce the early warning frequency.
[0111] For the monitoring effect of the generator state of the target domain unit without migration conditions, see Figure 5 (a). For the monitoring effect of the power generation state of the target domain unit using the method of the present invention, see Figure 5 (b). Using the method of the present invention can effectively monitor whether the operation state of the generator of the wind turbine unit is normal, and has a low early warning frequency, which can improve the operation and maintenance efficiency of the wind farm.
[0112] The embodiments described above are only used to describe the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for monitoring the state of a wind turbine based on a GAN-QP feature migration model, characterized in that, It includes the following steps: S1. Collect the historical dataset of the operating status of the generator systems of two wind turbines with the same unit type but different geographical locations and rated powers; S2. Take one of the wind turbines as the source domain unit and the other wind turbine as the target domain unit, and preprocess the historical datasets of the operating status of the generator systems of the two wind turbines respectively to obtain the source domain training set and the target domain training set; S3. Train the feature transfer model based on the target domain training set and the source domain training set to obtain the trained feature transfer model, and the trained feature transfer model can transform the feature distribution of the operating data of the target domain unit to the feature distribution of the operating data of the source domain unit; S4. Train the autoencoder based on the source domain training set to learn the feature distribution rules of the data under the normal operating state of the source domain unit to obtain the trained autoencoder; S5. Preprocess the online dataset of the operating status of the generator system of the target domain unit to obtain the target domain test set, and transform the target domain test set to the source domain space based on the trained feature transfer model to obtain the source domain test set; S6. Use the source domain test set obtained in step S5 as the input data of the trained autoencoder, calculate the mean square error between the source domain test set and the source domain training set, and judge whether the operating status of the generator system is normal based on the mean square error; In step S3, it includes the following steps: S3.
1. Divide the target domain training set into multiple batches; S3.
2. Select a certain batch of the target domain training set and input it into the generator network for preprocessing to obtain the generated dataset; S3.
3. Input the generated dataset and the source domain training set into the discriminator network to output the discrimination results of the two; S3.
4. Calculate the loss function values of the generator network and the discriminator network based on the discrimination results, and update the model parameters of the generator network and the discriminator network based on the loss function values; S3.
5. Loop through steps S3.2 to S3.4 for a preset number of times so that the feature distribution of the generated dataset of the generator network approximates the feature distribution of the source domain training set, and use the finally obtained generator network as the trained feature transfer model; The loss function L adopted by the discriminator network D is as follows: Among them, D(·) represents the discriminator output, x g represents the data in the generated dataset, x r represents the data in the source domain training set, λ is the balance coefficient, d(x r , x g ) represents the Euclidean distance between x r and x g . p r is the data distribution of the source domain training set, p g is the data distribution of the generated dataset. is to calculate the expected value for each sample data in the data distribution of the source domain training set and the data distribution of the generated dataset.
2. The method for monitoring the state of a wind turbine based on the GAN-QP feature transfer model according to claim 1, wherein In step S2, when preprocessing the historical datasets of the operating status of the generator systems of the source domain unit and the target domain unit, it includes the following steps: S2.
1. Delete the abnormal data in the historical dataset; S2.
2. Normalize the historical dataset processed in step S2.
1.
3. The wind turbine condition monitoring method based on the GAN-QP feature migration model according to claim 2, wherein, In step S2.1, the abnormal data includes the data with missing values, the shutdown data, the data outside the normal operating wind speed range of the wind turbine, and the outlier data.
4. The method for monitoring the state of a wind turbine based on the GAN-QP feature transfer model according to claim 3, characterized in that The judgment of the outlier data includes the following steps: A. Select a certain data point in the historical dataset; B. Select multiple target data points in the historical dataset that are closest to the data point selected in step A; C. Calculate the average chain distance of the selected data point based on the data point selected in step A and the multiple target data points; D. Calculate the distance based on the connection anomaly factor of the selected data point based on the average chain distance of the selected data point and the average chain distances corresponding to the multiple target data points respectively; E. Determine whether the selected data point is an outlier based on the connection anomaly factor distance of the selected data point.
5. A method for monitoring the state of a wind turbine based on a GAN-QP feature migration model according to claim 4, characterized in that In step D, the following steps are included: Calculate the average chain distance corresponding to each of multiple target data points; Calculate the average value of the average chain distances of multiple target data points; Based on the average chain distance of the selected data point and the average value of the average chain distances of multiple target data points, calculate the connection anomaly factor distance of the selected data point.
6. The method for monitoring the state of a wind turbine based on a GAN-QP feature transfer model according to claim 5, characterized in that The average chain distance of the selected data point where k is the total number of target data points, and e i represents the distance from the i-th target data point to the selected data point.
7. A wind turbine condition monitoring method based on the GAN-QP feature migration model according to claim 6, characterized in that Connection Outlier Factor Distance COF of Selected Data Points k (p) is calculated as follows: Among them, N k (p) represents multiple sets of target data points, represents the average chain distance of the o-th target data point.
8. A method for monitoring the state of a wind turbine based on a GAN-QP feature transfer model according to claim 1, characterized in that, The loss function L adopted by the generator network G is as follows: