Wind turbine generator fault early warning method based on adaptive double control strategy

Abstract

The present application relates to the technical field of fault early warning, in particular to a wind turbine generator fault early warning method based on adaptive double control strategy, which comprises the following steps: S1, obtaining historical operation data of the wind turbine by using a data acquisition and monitoring control system, and preprocessing the historical operation data; S2, importing the data set into a training model, setting the input and output of the training model, training the model, and saving the trained model as a category gradient boosting benchmark model; S3, finding the optimal hyperparameter variable of the category gradient boosting benchmark model by using random search; S4, calculating the residual error between the predicted value and the actual value, and constructing an adaptive double control strategy to dynamically determine the change trend of the residual error; S5, determining the fault point by using the adaptive double control strategy. The present application can be used for real-time early warning and alarm of offshore wind turbine groups and land wind turbine groups, and has the advantages of robustness, universality, accuracy and high efficiency in fault early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault early warning, in particular to a wind turbine generator fault early warning method based on an adaptive double control strategy. BACKGROUND

[0002] Wind energy is one of the renewable energy sources with the most prospects for large-scale development and commercialization, and is also the fastest developing renewable energy source. However, the working environment of a wind turbine is very harsh, usually located in remote suburbs and harsh coastal areas, and operates all day long under alternating loads, resulting in frequent failures and causing huge maintenance costs and economic losses.

[0003] The most important component in a wind turbine is the generator, so for intelligent operation and maintenance of a wind turbine generator, SCADA (Supervisory Control And Data Acquisition) data are processed and analyzed, and most algorithms for early warning are related to components by using statistical indicators or setting hard threshold indicators. The setting of a hard threshold requires experience and has a high false alarm rate. Therefore, by analyzing SCADA using a big data algorithm, data features are extracted to increase the accuracy of the algorithm.

[0004] Traditional random forest, GBDT and XGBoost algorithms require adjustment of a large number of parameters and have high training time costs. The required algorithm is to process a large amount of continuous data of a wind turbine in batches, and the data volume is large and the dimension is high, so the universality of the traditional algorithm is not high. Therefore, an early warning algorithm with strong adaptability, high stability and fast data processing is an essential part of current wind farm health monitoring. SUMMARY

[0005] Therefore, it is necessary to provide a wind turbine generator fault early warning method based on an adaptive double control strategy in view of the above technical problems.

[0006] The present application provides a wind turbine generator fault early warning method based on an adaptive double control strategy, which comprises the following steps:

[0007] S1, obtaining historical operation data of a wind turbine using a data acquisition and monitoring control system, and preprocessing the historical operation data to finally combine to form a new data set;

[0008] S2, importing the data set into a training model, setting the input and output of the training model, training the model, and saving the trained model as a category gradient boosting benchmark model;

[0009] S3, using random search to find the optimal hyperparameter variable of the category gradient boosting benchmark model;

[0010] S4, calculate the residual between the predicted value and the actual value, and re-construct an adaptive double control strategy to dynamically determine the trend of the residual to prevent false alarms and early alarms;

[0011] S5, in the process of executing new data prediction, pre-load the category gradient boosting benchmark model and the predicted data, determine the fault point through the adaptive double control strategy, and output the early warning information.

[0012] Further, the historical operation data is pre-processed, and finally combined to form a new data set including the following steps:

[0013] S11, the special point data in the historical operation data is removed;

[0014] S12, the data characteristics of the historical operation data are extracted by using the Pearson correlation coefficient algorithm, the data columns with a correlation value greater than 0.5 are selected, and the data columns are qualitatively deleted parameters;

[0015] S13, the finally retained data columns are combined to form a new data set.

[0016] Further, the special point data includes negative values, zero values, infinite values, null values and random codes, and the qualitative deletion parameters include the data of blades 2 and 3 in the wind turbine generator.

[0017] Further, the input and output of the training model are set, the model is trained, and after the training is completed, it is saved as a category gradient boosting benchmark model including the following steps:

[0018] S21, convert the data format of the data set to the data set format, and then use the max-min normalization algorithm to standardize the data in the data set;

[0019] S22, divide the data set into a training set, a test set and a validation set in a ratio of 7:2:1;

[0020] S23, set the training set as the input end of the training model, set the predicted temperature of the generator drive end bearing as the output end of the training model, select the default basic hyperparameters, and train the model. Finally, after the training is completed, it is saved as a category gradient boosting benchmark model.

[0021] Further, the optimal hyperparameter variable of the category gradient boosting benchmark model is found by using random search including the following steps:

[0022] S31, select the key parameters of the category gradient boosting benchmark model for setting and adjusting, select the default values for the remaining parameters, and check the fitting or underfitting before parameter adjustment;

[0023] S32, the optimal parameters are automatically searched by using a grid search method, the variables are arranged and combined, and each combination is traversed, a plurality of hyperparameter selections are verified on the training set and the verification set in a cross-validation manner, and the hyperparameters with the minimum average error are selected;

[0024] S33, after the hyperparameters are selected, the training set and the verification set are merged, and the category gradient boosting benchmark model is trained again to obtain a final category gradient boosting benchmark model;

[0025] S34, the category gradient boosting benchmark model is tested by using the test set.

[0026] Further, the key parameters include the number of iterations, the minimum number of trees that the optimal model should have, and the tree depth.

[0027] Further, the residual between the predicted value and the actual value is calculated, and an adaptive double control strategy is constructed to dynamically determine the trend of the residual to prevent false alarms and early alarms, including the following steps:

[0028] S41, the verification set is imported into the category gradient boosting benchmark model to obtain the predicted value of the generator driving end bearing temperature, the predicted value is converted to the same dimension as the temperature by using the inverse normalization method, and then the residual is calculated with the actual value of the generator driving end bearing temperature to obtain the residual curve;

[0029] S42, the residual is solved by hard threshold and soft threshold respectively, and the soft and hard double threshold protection of the residual value is set as an adaptive double control strategy to prevent false alarms and early alarms of the category gradient boosting benchmark model.

[0030] Further, the solving method of the hard threshold is to assume that a group of residual data only contains random errors, calculate the standard deviation of the original data, and then determine a judgment interval according to a preset probability. If the error exceeds the judgment interval, it is judged that the data belongs to an abnormal value.

[0031] Further, the solving method of the soft threshold is to compare whether two sample data come from the same distribution, dynamically calculate the error distribution of each interval, and the two sample data respectively represent the actual value and the predicted value of the generator driving end temperature.

[0032] Further, the fault point is judged by using the adaptive double control strategy, and the warning information is output, including the following steps:

[0033] S51, if the residual value single point is greater than the soft threshold curve, observe the residual point position;

[0034] S52, if the residual value single point is greater than the hard threshold curve, it is judged that the fault is serious and the shutdown inspection of the wind turbine generator needs to be performed;

[0035] S53, if the residual value single point is less than the soft threshold curve, it is judged that the time is normal;

[0036] S54, if the residual data single point is greater than the soft threshold curve and less than the hard threshold curve, it is judged that a fault is about to occur, and the wind turbine is in an observed state during the alarm process.

[0037] The beneficial effects of the present application are: by providing a wind turbine generator fault early warning method based on adaptive double control strategy, the redundancy of data can be prevented, and in the data cleaning process, the historical operation data is integrated and reduced in dimension by using double feature extraction; in view of the problems of large amount of data and high dimension of SCADA monitoring and collection, the efficient characteristics of CatBoost are used to reduce the selection of basic parameter amount, and the time cost problem caused by large amount of data is well solved; in view of the alarm event, it is divided into four parts of normal, early warning, time observation and shutdown inspection, which is more reasonable to reserve time for fault occurrence and control the economic cost caused by advance shutdown; the present application can be used for real-time early warning and alarm of offshore wind turbine group and land wind turbine group, and has robustness, universality, accuracy and high efficiency in fault early warning. BRIEF DESCRIPTION OF DRAWINGS

[0038] The drawings described herein are used to provide further understanding of the present application, and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute improper limitation on the present application. In the drawings:

[0039] Figure 1 It is a flow chart of a wind turbine generator fault early warning method based on adaptive double control strategy according to an embodiment of the present application;

[0040] Figure 2 It is a fault early warning flow chart of a wind turbine generator fault early warning method based on adaptive double control strategy according to an embodiment of the present application;

[0041] Figure 3 It is a temperature regression prediction algorithm flow chart of a wind turbine generator fault early warning method based on adaptive double control strategy according to an embodiment of the present application;

[0042] Figure 4 It is a generator driving end bearing temperature prediction diagram of a wind turbine generator fault early warning method based on adaptive double control strategy according to an embodiment of the present application;

[0043] Figure 5 It is one of early warning effect diagrams obtained by adaptive double control strategy of a wind turbine generator fault early warning method based on adaptive double control strategy according to an embodiment of the present application;

[0044] Figure 6is the second early warning effect diagram obtained by the adaptive double control strategy of the wind turbine generator fault early warning method based on the adaptive double control strategy according to the embodiment of the application;

[0045] Figure 7 is the third early warning effect diagram obtained by the adaptive double control strategy of the wind turbine generator fault early warning method based on the adaptive double control strategy according to the embodiment of the application;

[0046] Figure 8 is the fourth early warning effect diagram obtained by the adaptive double control strategy of the wind turbine generator fault early warning method based on the adaptive double control strategy according to the embodiment of the application;

[0047] Figure 9 is the fifth early warning effect diagram obtained by the adaptive double control strategy of the wind turbine generator fault early warning method based on the adaptive double control strategy according to the embodiment of the application. DETAILED DESCRIPTION

[0048] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0049] Please refer to Figures 1-3 , which provides a wind turbine generator fault early warning method based on an adaptive double control strategy, which comprises the following steps:

[0050] S1, obtaining historical operation data of a wind turbine by using a data acquisition and monitoring control system, and pre-processing the historical operation data to finally combine to form a new data set.

[0051] The data acquisition and monitoring control system (Supervisory Control And Data Acquisition, SCADA) is a system used for real-time monitoring and control of industrial processes.

[0052] The SCADA system usually contains the following key components: remote terminal unit (RTU) or PLC, sensor and actuator, communication network, SCADA host, human-machine interface (HMI), database system, alarm system and remote access, and the SCADA system is widely used in industrial, energy, public utilities and other fields, which provides real-time monitoring, fault diagnosis and remote control capabilities for industrial processes, and helps to improve production efficiency and system reliability.

[0053] In the description of the present application, the pre-processing of the historical operation data to finally combine to form a new data set comprises the following steps:

[0054] S11, special point data in the historical operation data is removed.

[0055] The special point data includes negative value, zero value, infinite value, null value and garbled code.

[0056] S12, the data characteristics of the historical operation data are extracted by using the Pearson correlation coefficient algorithm, the data column with a correlation value greater than 0.5 is selected, and the data column is qualitatively deleted.

[0057] The Pearson correlation coefficient algorithm has the characteristics of good noise robustness and sensitive correlation change degree, the data characteristics are extracted, the correlation of the data is analyzed, and the data column with a correlation value greater than 0.5 is extracted, and the extracted data column is qualitatively deleted, which helps to reduce data redundancy.

[0058] The deleted data includes the data of paddle 2 and paddle 3, because the three paddles have the same effect on the generator.

[0059] S13, the finally reserved data column is combined to form a new data set.

[0060] S2, the data set is imported into the training model, the input and output of the training model are set, the model is trained, and the trained model is saved as a category gradient boosting (CatBoost) benchmark model.

[0061] In the description of the application, the input and output of the training model are set, the model is trained, and the trained model is saved as a category gradient boosting benchmark model, including the following steps:

[0062] S21, the data format of the data set is converted into a data set (Dataset) format, and the data in the data set is standardized by using the max-min normalization algorithm.

[0063] The max-min normalization algorithm is integrated in the function MinMaxScaler () in the sklearn library (the library is a machine learning algorithm library); wherein the max-min normalization algorithm scales the features between the given minimum value and the maximum value, and can also convert the maximum absolute value of each feature to a unit size. This method is a linear transformation of the original data, which normalizes the data to [0, 1], and can ensure that the abnormal value data points also show regularity within a certain range.

[0064] S22, the data set is divided into a training set, a test set and a validation set in the ratio of 7:2:1.

[0065] S23, set the training set as the input of the training model, set the predicted temperature of the generator driving end bearing as the output of the training model, and then select default basic hyperparameters, perform model training, and finally save the trained model as a CatBoost benchmark model.

[0066] CatBoost is a machine learning framework based on gradient boosting algorithm, which is specially designed for processing categorical features. It can achieve good performance on classification and regression problems, and also performs well on large-scale data sets. It can automatically process categorical features without one-hot encoding or other tedious preprocessing steps, making it more convenient for processing actual data sets. It implements gradient boosting-based algorithms that can effectively learn non-linear relationships and complex patterns, and usually performs well, especially on large-scale data sets. It can automatically handle missing values without additional user processing, making model training easier and more robust. In addition, CatBoost uses effective regularization techniques to prevent overfitting and improve model generalization.

[0067] S3, use random search to find the optimal hyperparameter variable of the CatBoost benchmark model.

[0068] In the description of the present application, the use of random search to find the optimal hyperparameter variable of the CatBoost benchmark model includes the following steps:

[0069] S31, select and adjust the key parameters of the CatBoost benchmark model, select the default values for the remaining parameters, and check the fitting or underfitting before parameter adjustment.

[0070] The key parameters include the number of iterations (iterations), the minimum number of trees that the best model should have (best_model_min_trees), and the tree depth (depth).

[0071] S32, use the grid search method to automatically search for the optimal parameters, arrange and combine the variables, and traverse each combination. Through cross-validation, the average error of the selected hyperparameters is verified on the training set and the validation set, and the hyperparameters with the smallest average error are selected.

[0072] S33, after selecting the hyperparameters, merge the training set and the validation set, and then retrain the CatBoost benchmark model to obtain the final CatBoost benchmark model.

[0073] S34, test the CatBoost benchmark model using the test set.

[0074] S4. Calculate the residual between the predicted value and the actual value, and then construct an adaptive dual control strategy to dynamically judge the changing trend of the residual and prevent false alarms and early warnings.

[0075] In this case, the cross-validation parameter cv is set to 3, the model evaluation criterion scoring is set to r2, and the other parameters are set to their default values.

[0076] In the description of this invention, the process of calculating the residual between the predicted value and the actual value, constructing an adaptive dual control strategy, and dynamically judging the changing trend of the residual to prevent false alarms and early warnings includes the following steps:

[0077] S41. Import the validation set into the class gradient boosting baseline model to obtain the predicted value of the generator drive end bearing temperature. Convert the predicted value into a dimension consistent with temperature using the inverse normalization method, and then perform residual calculation with the actual value of the generator drive end bearing temperature to obtain the residual curve.

[0078] S42. Solve for the hard threshold (3sigama hard threshold) and soft threshold (KS soft threshold) of the residual respectively. Set the soft and hard double-layer threshold protection of the residual value as an adaptive dual control strategy to prevent false alarms and early warnings in the class gradient boosting benchmark model.

[0079] The solution for the hard threshold is to assume that a set of residual data contains only random errors, calculate and process the original data to obtain the standard deviation, and then determine a judgment interval according to a preset probability. If the error exceeds the judgment interval, the data is judged to be an outlier.

[0080] The soft threshold solution involves comparing whether two sample data points come from the same distribution, dynamically calculating the error distribution for each interval, and having the two sample data points represent the actual and predicted values ​​of the generator drive end temperature in separate tables.

[0081] By using both hardware and software threshold protection, false alarms caused by the model are prevented. Therefore, the designed dual control strategy greatly improves the accuracy of alarms.

[0082] S5. During the new data prediction process, the category gradient boosting baseline model and predicted data are preloaded, and the fault point is identified through an adaptive dual control strategy, and early warning information is output.

[0083] In the description of this invention, determining the fault point and outputting early warning information through an adaptive dual-control strategy includes the following steps:

[0084] S51. If the residual value at a single point is greater than the soft threshold curve, then carefully observe the location of the residual point.

[0085] S52, if the residual value single point is greater than the hard threshold curve, it is judged that the fault is large, and shutdown inspection of the wind turbine needs to be performed.

[0086] S53, if the residual value single point is less than the soft threshold curve, it is judged that the time is normal.

[0087] S54, if the residual data single point is greater than the soft threshold curve and less than the hard threshold curve, it is judged that the fault is about to occur, and the wind turbine is in an observed state during the alarm process.

[0088] As shown in the wind turbine generator drive end bearing temperature prediction effect diagram, Figure 4 the GridSearchCV is used to automatically search for the optimal parameters, find the optimal solution for the data set prediction, and use the RMS and R2 as evaluation indexes, as shown in the prediction effect. Figure 4

[0089] As shown in the early warning effect diagram obtained by using the adaptive double control strategy proposed in the application, Figures 5-9 it can be clearly seen that the application can not only provide early warning but also prevent false positives.

[0090] To sum up, by means of the above technical scheme of the application, through the wind turbine generator fault early warning method of adaptive double control strategy, the redundancy of data can be prevented, and in the data cleaning process, the historical operation data is integrated and reduced in dimension by using double feature extraction; in view of the problems of large amount of data and high dimension of SCADA monitoring and collection, the CatBoost high efficiency is used to reduce the selection of basic parameter amount, and the time cost problem caused by large amount of data is well solved; in view of the alarm event, it is divided into four parts: normal, early warning, observation, and shutdown inspection, which can more reasonably reserve time for fault occurrence and control the economic cost caused by early shutdown; the application can be used for real-time early warning and alarm of offshore wind turbine groups and land wind turbine groups, and has robustness, universality, and accurate and efficient fault early warning.

[0091] It should be understood that although each step in the flowchart of the accompanying drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other sequences. Moreover, at least part of the steps in the flowchart of the accompanying drawings can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or other steps, sub-steps or stages.​

Claims

1. A method for early warning of generator faults in wind turbines based on an adaptive dual control strategy, characterized in that, include: S1. Use the data acquisition and monitoring control system to acquire historical operating data of the wind turbine, preprocess the historical operating data, and finally combine them to form a new dataset; S2. Import the dataset into the training model, set the input and output of the training model, train the model, and save it as a class gradient boosting baseline model after training is complete. S3. Use random search to find the optimal hyperparameter variables of the class gradient boosting baseline model; S4. Calculate the residual between the predicted value and the actual value, and then construct an adaptive dual control strategy to dynamically judge the changing trend of the residual and prevent false alarms and early warnings. S5. During the new data prediction process, the category gradient boosting benchmark model and the predicted data are preloaded, and the fault point is determined by the adaptive dual control strategy, and early warning information is output. The process involves calculating the residual between the predicted and actual values, then constructing an adaptive dual-control strategy to dynamically determine the trend of the residual's change, preventing false alarms and early warnings. S41. Import the validation set into the class gradient boosting baseline model to obtain the predicted value of the generator drive end bearing temperature. Convert the predicted value into a dimension consistent with temperature using the inverse normalization method. Then perform residual calculation with the actual value of the generator drive end bearing temperature to obtain the residual curve. S42. Solve for hard and soft thresholds for the residuals respectively, and set the hard and soft double-layer threshold protection of the residual values ​​as an adaptive dual control strategy to prevent false alarms and early warnings in the class gradient boosting benchmark model. The method of determining the fault point through adaptive dual control strategy and outputting early warning information includes: S51. If the residual value at a single point is greater than the soft threshold curve, then carefully observe the location of the residual point. S52. If the residual value at a single point is greater than the hard threshold curve, it is determined that the fault is serious and the wind turbine needs to be shut down for inspection. S53. If the residual value at a single point is less than the soft threshold curve, then the judgment time is normal. S54. If a single point of the residual data is greater than the soft threshold curve but less than the hard threshold curve, it is determined that a fault is about to occur, and the wind turbine is under observation during the alarm process.

2. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 1, characterized in that, The process of preprocessing the historical operational data and ultimately combining it to form a new dataset includes the following steps: S11. Remove special point data from the historical operation data; S12. The Pearson correlation coefficient algorithm is used to extract the data features of the historical operation data, select the data column with a correlation value greater than 0.5, and qualitatively delete parameters from the data column; S13. Combine the data columns that are ultimately retained to form a new dataset.

3. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 2, characterized in that, The special point data includes negative values, zero values, infinite values, null values, and garbled characters. The qualitative deletion parameters include the data of blade 2 and blade 3 in the wind turbine.

4. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 1, characterized in that, The process of setting the input and output of the training model, training the model, and saving it as a class gradient boosting baseline model after training includes the following steps: S21. Convert the data format of the dataset to the dataset format, and then use the minimax normalization algorithm to perform data standardization on the data in the dataset. S22. Divide the dataset into a training set, a test set, and a validation set in a ratio of 7:2:1; S23. Set the training set as the input of the training model, set the predicted temperature of the generator drive bearing as the output of the training model, select the default basic hyperparameters, train the model, and finally save it as a class gradient boosting baseline model after training is completed.

5. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 4, characterized in that, The process of finding the optimal hyperparameter variables for the class gradient boosting baseline model using random search includes the following steps: S31. Select the key parameters of the category gradient boosting benchmark model and set and adjust them. Select the default values ​​for the remaining parameters and check for fitting or underfitting before adjusting the parameters. S32. The grid search method is used to automatically search for the optimal parameters. The variables are arranged and combined, and each combination is traversed. The selection of multiple hyperparameters is verified on the training set and the validation set through cross-validation, and the hyperparameter with the smallest average error is selected. S33. After the hyperparameters are selected, the training set and the validation set are merged, and the class gradient boosting benchmark model is trained again to obtain the final class gradient boosting benchmark model. S34. Test the category gradient boosting benchmark model using the test set.

6. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 5, characterized in that, The key parameters include the number of iterations, the minimum number of trees that the optimal model should have, and the tree depth.

7. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 6, characterized in that, The method for solving the hard threshold is to assume that a set of residual data contains only random errors, calculate and process the original data to obtain the standard deviation, and then determine a judgment interval according to a preset probability. If the error exceeds the judgment interval, the data is judged to be an outlier.

8. The wind turbine generator fault early warning method based on adaptive dual control strategy according to claim 7, characterized in that, The soft threshold is calculated by comparing whether two sample data come from the same distribution and dynamically calculating the error distribution of each interval. The two sample data represent the actual value and the predicted value of the generator drive end temperature, respectively.

Images