Unit operation data processing method for temperature fault early warning of large components of wind turbine generator

Abstract

The invention provides a unit operation data processing method for temperature fault early warning of large components of a wind turbine generator. The method comprises the steps of drawing a wind speed and power scatter diagram according to operation data of a wind turbine generator, dividing a plurality of different types of abnormal temperature points in the wind speed and power scatter diagram, and removing the different types of abnormal temperature points by adopting a plurality of different methods to obtain cleaned data samples; each piece of data of the cleaned data sample comprises a power value and a plurality of feature parameters, calculating each feature parameter by adopting at least two different feature extraction methods to obtain a correlation degree value, and retaining the corresponding feature parameter or deleting the corresponding feature parameter according to each correlation degree value, the reserved characteristic parameters and the power values form a training sample, and the training sample is used for training a temperature prediction model. According to the method, the cleaning efficiency and effect can be improved, and feature parameters with higher relevancy can be screened out.

Description

Technical Field

[0001] The present application belongs to the technical field of temperature fault warning for large components of wind turbines, and specifically relates to a method for processing unit operation data for temperature fault warning for large components of wind turbines. Background Art

[0002] In order to obtain high-quality wind energy resources, wind turbines are usually located in remote areas with high wind speeds, such as mountains, coasts, and Gobi Deserts. This makes the data transmission system prone to failure, resulting in partial data loss. At the same time, the turbine will generate abnormal data during maintenance, operating mode switching, wind curtailment and power rationing, and fault conditions. In addition, due to the limitations of sensor measurement frequency and accuracy, a large amount of duplicate data will appear in certain periods. Since these abnormal data cannot reflect the normal operating rules of the turbine, they will have a significant negative impact on the training of the temperature prediction model under normal operating conditions used for temperature fault warning of large components of wind turbines. Therefore, it is necessary to clean the turbine data to provide a reliable data source for model training, thereby avoiding the decline of model performance and warning effect. However, the existing abnormal data cleaning methods have many shortcomings: (1) Over-reliance on manual experience to judge abnormal operating conditions. When dealing with complex turbine operating conditions, the data elimination process takes a long time. (2) Most of them use a single supervised learning or clustering algorithm for data cleaning, and lack comparative analysis of the effects of different algorithms. (3) The parameters of the supervised learning or clustering algorithm are often set by default, and the cleaning effects of different parameter combinations are not evaluated. (4) For accumulated data close to the normal area, the cleaning effect is poor.

[0003] Furthermore, given the numerous measurement points and diverse monitoring objectives within a unit, model training should select the variables most relevant to the monitoring objective. This reduces data dimensionality while retaining key information, thereby improving model performance. However, existing feature extraction methods also have shortcomings: over-reliance on a single feature extraction method makes them susceptible to inherent flaws. For example, some filtering methods fail to account for nonlinear relationships between variables. Summary of the Invention

[0004] The technical problem solved by this application is: how to improve the abnormal data cleaning effect of the unit operation data for temperature fault warning of large components of wind turbines and extract more effective features.

[0005] The present application provides a method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine, the method comprising:

[0006] Data cleaning step: drawing a wind speed and power scatter plot based on the wind turbine operating data, dividing several different types of abnormal temperature points in the wind speed and power scatter plot, and using several different methods to remove the different types of abnormal temperature points respectively to obtain cleaned data samples;

[0007] Feature screening step: Each piece of data in the cleaned data sample includes a power value and several characteristic parameters. For each of the characteristic parameters, at least two different feature extraction methods are used to calculate the correlation degree value. According to each correlation degree value, the corresponding characteristic parameter is retained or deleted. The retained characteristic parameters and power values constitute training samples, and the training samples are used to train the temperature prediction model.

[0008] Optionally, the method of dividing a plurality of different types of abnormal temperature points in the wind speed and power scatter diagram includes:

[0009] The first type of abnormal temperature point is the data in the form of a band that is located near zero power and extends vertically to the left and right on the scatter plot.

[0010] The data on the scatter plot that are far away from the normal wind speed-power curve and distributed in the medium-low power area and the medium-high power area with a wind speed of 10m / s to 25m / s are regarded as the second type of abnormal temperature points;

[0011] On the scatter plot, the third type of abnormal temperature point is located in the middle power area of wind speed 10m / s to 25m / s and is represented by multiple point bands extending horizontally to the right.

[0012] Optionally, the method of using several different methods to remove different types of abnormal temperature points to obtain cleaned data samples includes:

[0013] For the first type of abnormal temperature points, the temperature points with power less than or equal to 0 are deleted, the missing temperature points are filled, and the duplicate temperature points are deleted.

[0014] Optionally, the method of using several different methods to remove different types of abnormal temperature points to obtain cleaned data samples includes:

[0015] The second type of abnormal temperature points and a portion of the third type of abnormal temperature points are identified and removed using the KNN algorithm. After removal, the other portion of the third type of abnormal temperature points are located near the main band of the low power area in the scatter plot.

[0016] Optionally, the method of using several different methods to remove different types of abnormal temperature points to obtain cleaned data samples further includes:

[0017] Construct normal operating intervals in the scatter plot;

[0018] The abnormal temperature points of the third type located outside the normal operating range are removed.

[0019] Optionally, the method of constructing a normal operating interval in a scatter plot includes:

[0020] Divide the remaining data of the scatter plot into several subintervals and determine the power median point of each subinterval;

[0021] The wind speed-power curve under the actual state of the unit is obtained by fitting the power median point of each sub-interval;

[0022] Use the fitted curve to divide each subinterval into two parts, left and right, and determine the upper and lower power points of each subinterval;

[0023] An upper limit curve is obtained by fitting the upper limit power point of each sub-interval, and a lower limit curve is obtained by fitting the lower limit power point of each sub-interval. The normal operating range is between the upper limit curve and the lower limit curve.

[0024] Optionally, the method of respectively calculating the correlation degree value using at least two different feature extraction methods for each feature parameter includes:

[0025] Calculate the Spearman correlation coefficient of the characteristic parameter as the first type of correlation degree value;

[0026] The NMI value of the characteristic parameter is calculated as the second type of correlation degree value.

[0027] The present application provides a method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine, which has the following technical effects:

[0028] This method combines the advantages of integrated models and deep learning models in time series prediction, makes up for the inherent defects of a single model, and improves prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 The present invention is a flowchart of a method for processing wind turbine operation data for early warning of temperature failure of major components of a wind turbine according to one or more embodiments.

[0030] Figure 2 1 is a distribution diagram of abnormal temperature points of different types according to one or more embodiments.

[0031] Figure 3 FIG. 4 is a diagram showing the identification results of abnormal temperature points of the first type according to one or more embodiments.

[0032] Figure 4FIG. 4 is a schematic diagram of removing abnormal temperature points of the first type according to one or more embodiments.

[0033] Figure 5a 、 Figure 5b 、 Figure 5c Graph showing the recognition results of the KNN algorithm under different thresholds P according to one or more embodiments.

[0034] Figure 6 The figure is a wind speed-power scatter plot identified and processed by the optimal P KNN algorithm according to one or more embodiments.

[0035] Figure 7 The wind speed-power curve under actual conditions is fitted using the least squares method according to one or more embodiments.

[0036] Figure 8 Graphs showing the normal operating range and identification results obtained by fitting using the least squares method according to one or more embodiments.

[0037] Figure 9 FIG. 1 is a wind speed-power scatter plot of data after final cleaning is completed according to one or more embodiments. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0039] Before describing the various embodiments of the present application in detail, the technical concept of the present application will be briefly described first: When processing the unit operation data for temperature fault warning of large components of wind turbines, when cleaning the data, either relying on manual experience will lead to too low efficiency, or using a single cleaning method, and the cleaning effect is poor; when screening features, it is difficult to obtain effective feature parameters by using a single screening method. To this end, the present application provides a unit operation data processing method for temperature fault warning of large components of wind turbines. First, different types of temperature anomalies are divided, and different cleaning methods are used for targeted cleaning to improve cleaning efficiency and effect. Then, at least two methods are used to calculate the degree of correlation of feature parameters, and a comprehensive evaluation is performed to screen out feature parameters with higher correlation to form training samples. The specific principles of the unit operation data processing method for temperature fault warning of large components of wind turbines of the present application will be described below in combination with more embodiments.

[0040] Specifically, if Figure 1 The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine in the first embodiment includes the following steps:

[0041] Data cleaning step S10: drawing a wind speed and power scatter plot based on the wind turbine operating data, dividing several different types of abnormal temperature points in the wind speed and power scatter plot, and using several different methods to remove the different types of abnormal temperature points respectively to obtain cleaned data samples;

[0042] Feature screening step S20: Each piece of data in the cleaned data sample includes a power value and several characteristic parameters. For each characteristic parameter, at least two different feature extraction methods are used to calculate the correlation degree value. According to each correlation degree value, the corresponding characteristic parameter is retained or deleted. The retained characteristic parameters and power values constitute training samples, and the training samples are used to train the temperature prediction model.

[0043] For example, wind turbine operating data is first exported from the SCADA platform. Each piece of wind turbine operating data includes power values and characteristic parameters such as wind turbine wind speed and various temperatures. Some wind turbine operating data is shown in the following table.

[0044]

[0045] Table 1 Part of the original data

[0046] In one or more embodiments, a method for classifying several different types of abnormal temperature points in a wind speed and power scatter plot includes: selecting strip-shaped stacked data near zero power on the scatter plot and extending vertically left and right as a first type of abnormal temperature point, which is mainly caused by unit shutdown, unit sensor failure, and intentional wind curtailment, such as Figure 2 The scatter plot shows data that is far away from the normal wind speed-power curve and is distributed in the medium-low power area and the medium-high power area with a wind speed of 10m / s to 25m / s. As the second type of abnormal temperature points, there are fewer points in the same range and the point distribution is relatively random. It is usually caused by operating mode switching, extreme operating conditions, unit sensor failure and data transmission abnormality, such as Figure 2 As shown by the yellow circle in the figure. On the scatter plot, the third type of abnormal temperature point is in the medium power area of wind speed 10m / s to 25m / s and is represented by multiple horizontal point bands extending to the right. It is usually caused by the control unit operating at a fixed power, as shown in Figure 2 As shown in the red circle.

[0047] In one or more embodiments, for the first type of abnormal temperature points, temperature points with power less than or equal to 0 are deleted, missing temperature points are filled, and duplicate temperature points are deleted. Figure 3 Shows the abnormal temperature points that have been identified. Figure 4Almost all the banded data that showed zero power and extended vertically to the left and right were identified and removed. For example, the nearest neighbor interpolation method was used to fill in the missing data; adjacent data rows were compared, and if there were duplicates, only one record was retained.

[0048] In one or more embodiments, the KNN algorithm is used to identify and remove the second type of abnormal temperature points and a portion of the third type of abnormal temperature points. After removal, the other portion of the third type of abnormal temperature points are located near the main band of the low power area in the scatter plot. For example, the main process of abnormal data detection using the KNN algorithm includes:

[0049] (1) Extract data and map it into sample points in two-dimensional space.

[0050] (2) Calculate the Euclidean distance d of each sample point to all other sample points. The calculation formula is as follows:

[0051]

[0052] Among them, x1, x2, y1, and y2 are the horizontal and vertical coordinate values of any two points.

[0053] (3) Sort the distances calculated in (2) from small to large, and then determine the K other sample points closest to the sample point based on the set K value.

[0054] (4) Calculate the distance from each point to its Kth nearest neighbor and define it as the local density. The smaller the distance, that is, the larger the local density, the more likely this point is a normal value.

[0055] (5) Count the local densities of all points, arrange them from small to large, and set a threshold. When the local density of a point exceeds the threshold, the point can be considered an outlier.

[0056] For example, under the premise of setting K=10 and the distance metric criterion as Euclidean distance, the recognition effects of the KNN algorithm corresponding to different thresholds P are as follows: Figure 5a 、 Figure 5b 、 Figure 5c As shown in , when the threshold P becomes larger, the corresponding data classified as abnormal becomes less. Figure 5a As shown in Figure 1, when P = 88%, although the algorithm can identify most of the second and third types of data, it also misjudges some normal data on the outer edge as abnormal data, especially the right part. Figure 5c As shown in , when P = 98%, although the algorithm recognizes most of the normal data, it does not recognize most of the third type of data; Figure 5bAs shown in the figure, when P=93%, the algorithm recognizes the second type of data better, but there is still a small amount of third type of data that is not recognized. Therefore, when P is set to 93%, the KNN algorithm has the best recognition effect.

[0057] The wind speed-power scatter plot after identification and processing using the KNN algorithm under optimal P is as follows: Figure 6 As shown, it can be seen that after removing the identified abnormal data, some third type data still exists near the main band of the low power area in the scatter plot.

[0058] In one or more embodiments, for the remaining abnormal temperature points of the third type, a normal operating range is constructed in the scatter plot, and the abnormal temperature points of the third type outside the normal operating range are removed.

[0059] Exemplarily, the method for constructing a normal operating interval in a scatter plot includes the following steps:

[0060] (1) Divide the remaining data of the scatter plot into several subintervals and determine the power median point of each subinterval. Based on the data identified and cleaned by the KNN algorithm, the data is partitioned using fixed power intervals, and then the power values of all scattered points in each partition are arranged in ascending order, thereby calculating the median of each partition and the corresponding scattered point, which is defined as the power median point of the subinterval.

[0061] (2) The wind speed-power curve under the actual state of the unit is obtained by fitting the power midpoint of each sub-interval. For example, since the power is proportional to the cube of the wind speed, the least squares method is used to fit the power midpoint of each sub-interval with a cubic polynomial to draw the wind speed-power curve under the actual state of the unit, such as Figure 7 shown.

[0062] (3) Use the fitted curve to divide each subinterval into two parts, left and right, and determine the upper and lower power points of each subinterval. For example, according to the principle of the quartile algorithm, the wind speed value corresponding to the 50% power point on the left side of the subinterval is used as the lower quartile Q1, and the wind speed value corresponding to the 50% power point on the right side is used as the upper quartile Q3. Then, within the same power interval, the normal wind speed range is:

[0063] [Q1-1.5(Q3-Q1),Q3+1.5(Q3-Q1)]

[0064] (4) An upper limit curve is obtained by fitting the upper limit power points of each subinterval, and a lower limit curve is obtained by fitting the lower limit power points of each subinterval. The range between the upper limit curve and the lower limit curve is the normal operating range. For example, a least squares method is used to fit a cubic polynomial to the upper limit power point set and the lower limit power point set. Figure 8 The upper limit curve, lower limit curve, normal operating range and identified abnormal temperature points are shown. Figure 9 It is shown that almost all the remaining abnormal temperature points of the third type are excluded from the normal operating range and cleared.

[0065] In one or more embodiments, the method of calculating the correlation degree value for each of the feature parameters using at least two different feature extraction methods includes: calculating the Spearman correlation coefficient of the feature parameter as the first type of correlation degree value; calculating the NMI value of the feature parameter as the second type of correlation degree value.

[0066] For example, the Spearman correlation coefficient is calculated as follows:

[0067] (1) Convert the two variable data into rankings according to their size;

[0068] (2) Calculate the difference between the rankings of two variables in the same data and square it;

[0069] (3) Substitute the squares of all differences into the following formula to calculate the coefficients.

[0070]

[0071] Where ρ is the Spearman correlation coefficient, d i is the ranking difference of the i-th sample, and n is the number of samples.

[0072] For example, the calculation process of the NMI value is as follows:

[0073] First calculate MI:

[0074]

[0075] In order to better capture features with less information but high relevance to the target, MI(X, Y) is normalized by unilateral maximum entropy, which is calculated as follows:

[0076]

[0077] Where H(I) is the Shannon entropy, which is calculated as follows:

[0078]

[0079] Among them, p(I i ) is event I i Probability of occurrence.

[0080] For example, taking the gearbox bearing temperature as an example, the Spearman correlation coefficient and MI are used for feature selection, and the final Spearman and NMI values are shown in Table 2 below:

[0081]

[0082]

[0083] Table 2 Spearman coefficients and NMI values of relevant parameters

[0084] In Table 2, the gearbox oil sump temperature, generator stator U-phase coil temperature, generator stator V-phase coil temperature, and generator stator W-phase coil temperature all have large Spearman coefficients and NMI values, so they are included as parameters required for model training. However, since temperature signals are generally output variables, using output signals as the majority of variables in model training may lead to overfitting. Therefore, considering the physical space of the variables and the mechanical structure, generator active power, wind speed, generator speed, and rotor speed, which have large Spearman coefficients but small NMI values, can be included as parameters required for model training. Cabin temperature, cabin temperature, propeller motor temperature, propeller motor temperature, propeller motor temperature, propeller motor temperature, and propeller motor temperature, which have small Spearman coefficients but large NMI values, should be excluded. Grid-side frequency, grid-side L1 phase voltage, grid-side L2 phase voltage, and grid-side L3 phase voltage, which have both small Spearman coefficients and small NMI values, should also be excluded. Therefore, using this method, the final model input parameters determined are: generator active power, rotor speed, gearbox oil sump temperature, wind speed, generator speed, generator stator U-phase coil temperature, generator stator V-phase coil temperature, and generator stator W-phase coil temperature. Finally, the retained characteristic parameters and power values constitute the training samples used to train the temperature prediction model. The derived data serves as the input data for training the temperature prediction model under normal operating conditions.

[0085] The above describes in detail the specific implementation methods of the present application. Although some embodiments have been shown and described, those skilled in the art should understand that these embodiments can be modified and improved without departing from the principles and spirit of the present application, the scope of which is defined by the claims and their equivalents. These modifications and improvements should also be within the scope of protection of the present application.

Claims

1. A method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine, characterized in that: The method comprises: Data cleaning step: drawing a wind speed and power scatter plot based on the wind turbine operating data, dividing several different types of abnormal temperature points in the wind speed and power scatter plot, and using several different methods to remove the different types of abnormal temperature points respectively to obtain cleaned data samples; Feature screening step: Each piece of data in the cleaned data sample includes a power value and several characteristic parameters. For each of the characteristic parameters, at least two different feature extraction methods are used to calculate the correlation degree value. According to each correlation degree value, the corresponding characteristic parameter is retained or deleted. The retained characteristic parameters and power values constitute training samples, and the training samples are used to train the temperature prediction model.

2. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 1, characterized in that: The method for dividing several different types of abnormal temperature points in the wind speed and power scatter diagram includes: The first type of abnormal temperature point is the data in the form of a band that is located near zero power and extends vertically to the left and right on the scatter plot. The data on the scatter plot that are far away from the normal wind speed-power curve and distributed in the medium-low power area and the medium-high power area with a wind speed of 10m / s to 25m / s are regarded as the second type of abnormal temperature points; On the scatter plot, the third type of abnormal temperature point is located in the middle power area of wind speed 10m / s to 25m / s and is represented by multiple point bands extending horizontally to the right.

3. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 2, characterized in that: The method of using several different methods to remove different types of abnormal temperature points and obtain cleaned data samples includes: For the first type of abnormal temperature points, the temperature points with power less than or equal to 0 are deleted, the missing temperature points are filled, and the duplicate temperature points are deleted.

4. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 2, characterized in that: The method of using several different methods to remove different types of abnormal temperature points and obtain cleaned data samples includes: The second type of abnormal temperature points and a portion of the third type of abnormal temperature points are identified and removed using the KNN algorithm. After removal, the other portion of the third type of abnormal temperature points are located near the main band of the low power area in the scatter plot.

5. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 4, characterized in that: The method of using several different methods to remove different types of abnormal temperature points and obtain cleaned data samples also includes: Construct normal operating intervals in the scatter plot; The abnormal temperature points of the third type located outside the normal operating range are removed.

6. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 5, characterized in that: The method for constructing a normal operating interval in a scatter plot includes: Divide the remaining data of the scatter plot into several subintervals and determine the power median point of each subinterval; The wind speed-power curve under the actual state of the unit is obtained by fitting the power median point of each sub-interval; Use the fitted curve to divide each subinterval into two parts, left and right, and determine the upper and lower power points of each subinterval; An upper limit curve is obtained by fitting the upper limit power point of each sub-interval, and a lower limit curve is obtained by fitting the lower limit power point of each sub-interval. The normal operating range is between the upper limit curve and the lower limit curve.

7. The method for processing wind turbine operation data for early warning of temperature failure of large components of a wind turbine according to claim 1, characterized in that: The method of using at least two different feature extraction methods to calculate the correlation degree value for each feature parameter includes: Calculate the Spearman correlation coefficient of the characteristic parameter as the first type of correlation degree value; The NMI value of the characteristic parameter is calculated as the second type of correlation degree value.

Images