A fault warning modeling method and system for wind turbine generator sets

Abstract

The invention discloses a fault warning modeling method and system for a wind turbine generator set, the method comprising the following steps: S1, operation data acquisition; S2, label point data screening; S3, building a warning model; S4, determining a warning threshold; S5, judging the fault warning state of the wind turbine generator set; S6, testing the effect of the warning model; using the ReliefF algorithm to screen the SCADA data label points required for the warning model, and then using the machine learning algorithm to build the warning model, real-time warning of the wind turbine generator set components is performed to ensure that the model has a high accuracy rate, which is convenient for operation and maintenance personnel to timely troubleshoot faults; the invention is more scientific and reasonable in screening label point data; at the same time, the constructed warning model has a high accuracy rate, which improves the operation and maintenance efficiency; it has high compatibility and can be used for various key components of the wind turbine generator set, so as to timely discover the initial faults of the wind turbine generator set and avoid greater damage to the wind turbine.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine fault warning, and in particular to a wind turbine fault warning modeling method and system. Background Art

[0002] Domestic and foreign scholars' research on wind turbine fault warning is mostly achieved by using SCADA monitoring data to establish a normal wind turbine regression prediction model. The specific methods are as follows:

[0003] 1) Use neural network modeling, such as using neural network to calculate the Mahalanobis distance between the predicted value and the actual value to determine in advance whether the wind turbine is faulty.

[0004] 2) Use traditional machine learning modeling, such as SVM, to filter out the features of related components through algorithms, use SVM to predict the components, and finally realize fault warning by analyzing the residuals of model prediction.

[0005] 3) Use mathematical methods to establish regression models, such as establishing multiple regression analysis models.

[0006] 4) Use the continuity of time series to establish a model. For example, use the automatic regression sliding average model of the wind turbine vibration signal to identify the non-stationary state of the wind turbine during operation.

[0007] The above studies have achieved good results in wind turbine fault warning, but there are also some shortcomings. For example, neural networks require large-scale data when modeling, traditional machine learning often lacks accuracy and cannot achieve accurate prediction results, and time series prediction methods are generally suitable for short-term predictions, but have poor results in medium- and long-term predictions. Summary of the invention

[0008] The purpose of the present invention is to solve the deficiencies in the prior art and provide a fault warning modeling method and system for wind turbines. The ReliefF algorithm is used to screen the SCADA tag point data required for the warning model, and then the SVM machine learning algorithm is used to establish the warning model to provide real-time warnings for wind turbine components. By establishing a fault warning model, the key components of the wind turbine are warned, and the model is ensured to have a high accuracy rate, which is convenient for operation and maintenance personnel to troubleshoot faults in a timely manner and avoid greater damage to the wind turbine.

[0009] To achieve the above object, the present invention provides a technical solution: a fault warning modeling method for a wind turbine generator system, comprising the following steps:

[0010] S1. Collecting wind turbine operation data, preprocessing the collected wind turbine operation data, and dividing the data into a training sample set and a verification sample set;

[0011] S2, input the training sample set into the reliefF algorithm, obtain the feature influence value of each label point data, use the feature influence value as the weight of each label point data, and filter out each label point data according to the weight;

[0012] S3, input the filtered label point data into the SVM model for training and construct an early warning model;

[0013] S4, inputting the operating data of multiple groups of faulty wind turbines into the early warning model, obtaining the abnormal data proportions corresponding to the multiple groups of faulty wind turbines, and calculating the average value of all abnormal data proportions to obtain the early warning threshold;

[0014] S5. Inputting the wind turbine operating data currently required to be detected into the early warning model, comparing the abnormal data ratio of the wind turbine obtained with the early warning threshold, and determining whether a component of the wind turbine is faulty;

[0015] S6. Analyze the duration of each warning triggering according to the wind turbine operation data that triggers the warning to determine the severity of the wind turbine warning;

[0016] S7. Input the verification sample set into the early warning model to test the effectiveness of the early warning model.

[0017] Further, in step S1, the following operations are specifically performed:

[0018] The operation data of the wind turbine in normal operation and abnormal operation with a span of one month are obtained respectively, and the sampling period is 1s; the collected wind turbine data are preprocessed, and the first 80% of the collected wind turbine data are used as the training sample set, and the remaining 20% ​​of the data are used as the verification sample set.

[0019] Further, in step S3, the following operations are specifically performed:

[0020] The filtered label point data is input into the SVM model, and the SVM model is trained. The trained SVM model is the early warning model, and the early warning model is saved.

[0021] Further, in step S4, the following steps are included:

[0022] S401, after randomly taking the operation data of the faulty wind turbine set with a data length of 1 day and inputting it into the early warning model, the proportion of abnormal data judged as abnormal data in the operation data of the wind turbine set corresponding to that day is obtained;

[0023] S402, repeating step S401 multiple times, and taking the average value of all the abnormal data proportions obtained as the warning threshold.

[0024] Further, in step S5, the following operations are specifically performed:

[0025] After inputting the wind turbine operating data that needs to be detected currently into the early warning model, the abnormal data ratio of the wind turbine operating data is obtained. If the abnormal data ratio exceeds the early warning threshold, an early warning is triggered and the system prompts a unit component failure; if the abnormal data ratio does not exceed the early warning threshold, no early warning is triggered.

[0026] Further, in step S6, the following operations are specifically performed:

[0027] According to the operating data of the wind turbines that triggered the early warning, the duration of each early warning triggering is analyzed, and the severity of the wind turbine early warning is judged based on the duration; if the duration is less than one hour, it is judged as a level one warning; if the duration is greater than one hour and less than two hours, it is judged as a level two warning; if the duration is greater than two hours, it is judged as a level three warning.

[0028] The present invention provides a fault early warning modeling system for a wind turbine generator set, comprising:

[0029] An operation data acquisition module is used to acquire the operation data of the wind turbine generator set within a preset span time period;

[0030] Model the label acquisition module and filter out label point data through the ReliefF algorithm;

[0031] The model training module inputs the filtered corresponding label point data as a training set into the SVM model to build an early warning model;

[0032] A threshold determination module, used to determine a warning threshold based on wind turbine operation data;

[0033] The early warning module makes early warning judgments on the wind turbine operation data according to the early warning threshold, thereby issuing early warnings.

[0034] Furthermore, the operation data acquisition module specifically performs the following operations:

[0035] The operation data of the wind turbine in normal operation and abnormal operation with a span of one month are obtained respectively, and the sampling period is 1s; the collected wind turbine data are preprocessed, and the first 80% of the collected wind turbine data are used as the training sample set, and the remaining 20% ​​of the data are used as the verification sample set.

[0036] Furthermore, the threshold determination module specifically performs the following operations:

[0037] After randomly taking the operating data of a faulty wind turbine with a data length of 1 day and inputting it into the early warning model, the proportion of abnormal data judged as abnormal data in the corresponding wind turbine operating data of that day is obtained; repeating the above operation multiple times, the average value of all the abnormal data proportions obtained is used as the early warning threshold.

[0038] Furthermore, the early warning module specifically performs the following operations:

[0039] The abnormal data ratio of the wind turbine operating data is compared with the warning threshold. If the abnormal data ratio exceeds the warning threshold, the warning is triggered and the system prompts that the unit component is faulty; if the abnormal data ratio does not exceed the warning threshold, the warning is not triggered.

[0040] According to the operating data of the wind turbines that triggered the early warning, the duration of each early warning triggering is analyzed, and the severity of the wind turbine early warning is judged based on the duration; if the duration is less than one hour, it is judged as a level one warning; if the duration is greater than one hour and less than two hours, it is judged as a level two warning; if the duration is greater than two hours, it is judged as a level three warning.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0042] The present invention can determine the weight of each label point data through the reliefF algorithm, and filter the label point data according to the weight, so that the selection process is more scientific and reasonable; at the same time, the early warning model constructed by inputting the filtered label point data into the SVM model has a high accuracy rate, reduces invalid operation and maintenance, and improves operation and maintenance efficiency; it has high compatibility and can be used for various key components of the wind turbine generator set, so as to timely discover the initial faults of the wind turbine generator set and avoid greater damage to the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Flowchart of the fault warning modeling method for wind turbines.

[0044] Figure 2 A weight chart for each label point data.

[0045] Figure 3 It is a weighted line chart of each label point data.

[0046] Figure 4 This is a structural diagram for the training, calling and decision-making of the early warning model.

[0047] Figure 5 This is a chart of the warning thresholds for wind turbines within 15 days.

[0048] Figure 6 This is a statistical chart of the warning levels of wind turbines within 15 days.

[0049] Figure 7 This is the structural block diagram of the fault warning modeling system for wind turbines. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with specific embodiments.

[0051] See also Figure 1 As shown, the fault warning modeling method for wind turbines provided in this embodiment includes the following steps:

[0052] S1. Collect wind turbine operation data, pre-process the collected wind turbine operation data, and divide it into a training sample set and a verification sample set. Specifically, perform the following operations:

[0053] The operation data of the wind turbine blade motor in normal operation and abnormal operation with a span of one month are obtained respectively, and the sampling period is 1s; the collected wind turbine data are preprocessed, and the first 80% of the data is used as the training sample set, and the last 20% of the data is used as the verification sample set;

[0054] S2. Input the training sample set into the reliefF algorithm, obtain the feature influence value of each label point data, use the feature influence value as the weight of each label point data, and filter out each label point data according to the weight, where:

[0055] According to the SCADA tag point data of the wind turbine, the tag points related to the operating status of the wind turbine are as follows:

[0056]

[0057]

[0058] The above 20 tag point data are involved in total. According to the fault records, the operation data of the fault time period and the operation data of the normal operation time period are selected respectively, and the data types are marked;

[0059] See also Figures 2 to 3 As shown in the figure, the labeled data is input into the ReliefF algorithm model to obtain the weight of each label point data. According to the weight of each label point data, 8 label point data including operating status, generator power, blade 1 motor temperature, blade 2 motor temperature, blade 3 motor temperature, cabin temperature, blade angle and tower base temperature are selected from the 20 label point data for modeling.

[0060] S3. Input the filtered label point data into the SVM model for training and construct an early warning model, where:

[0061] See also Figure 4As shown, the 8 selected label point data are input into the SVM model for model training. The SVM classification model obtained after training is the early warning model, and the early warning model is saved. The model is called to classify each set of label point data that has been screened and tested for integrity, and then decide whether the wind turbine operation data related to the wind turbine and the blade motor at each moment is abnormal;

[0062] S4. Input the operating data of multiple groups of faulty wind turbines into the early warning model to obtain the abnormal data proportions corresponding to the multiple groups of faulty wind turbines, and calculate the average value of all abnormal data proportions to obtain the early warning threshold, where:

[0063] See also Figure 5 As shown, for the data of the wind turbine failure time period, the above-mentioned early warning model is called to calculate the abnormal data ratio of abnormal data in the daily wind turbine operation data; the average value of these abnormal data ratios is calculated as 9.333%, and the average value of 9.333% is used as the early warning threshold;

[0064] S5. Input the wind turbine operation data currently required to be detected into the early warning model, compare the abnormal data ratio of the wind turbine with the early warning threshold, and determine whether the component of the wind turbine is faulty. Specifically, perform the following operations:

[0065] After inputting the wind turbine blade motor operating data that needs to be detected into the early warning model, the abnormal data ratio of the wind turbine operating data is obtained. If the abnormal data ratio exceeds the early warning threshold of 9.333%, the blade motor fault early warning is triggered, and the system prompts a unit component failure; if the abnormal data ratio does not exceed the early warning threshold of 9.333%, the blade motor fault early warning is not triggered.

[0066] S6. According to the wind turbine operation data that triggers the early warning, analyze the duration of each early warning trigger to determine the severity of the wind turbine early warning, and perform the following operations:

[0067] See also Figure 6 As shown, according to the wind turbine operation data that triggers the warning, the duration of each warning is analyzed, and the severity of the wind turbine warning is judged according to the duration; if the duration is less than one hour, it is judged as a first-level warning; if the duration is greater than one hour and less than two hours, it is judged as a second-level warning; if the duration is greater than two hours, it is judged as a third-level warning;

[0068] S7. Input the verification sample set into the early warning model to test the effectiveness of the early warning model.

[0069] See also Figure 7 As shown, the following is a fault warning modeling system for a wind turbine provided in this embodiment, including:

[0070] The operation data acquisition module acquires the operation data of the wind turbine generator set within a preset span time period, and specifically performs the following operations:

[0071] The operation data of the wind turbine in normal operation and abnormal operation with a span of one month are obtained respectively, and the sampling period is 1s; the collected wind turbine data are preprocessed, and the first 80% of the collected wind turbine data are used as the training sample set, and the remaining 20% ​​of the data are used as the verification sample set.

[0072] Modeling label acquisition module, using ReliefF algorithm to filter out label point data associated with wind turbine components;

[0073] The model training module inputs the filtered corresponding label point data as a training set into the SVM model to build an early warning model;

[0074] The threshold determination module is used to determine the warning threshold based on the wind turbine operation data, and specifically performs the following operations:

[0075] After randomly taking the operating data of a faulty wind turbine with a data length of 1 day and inputting it into the early warning model, the proportion of abnormal data judged as abnormal data in the corresponding wind turbine operating data of that day is obtained; repeating the above operation multiple times, the average value of all the abnormal data proportions obtained is used as the early warning threshold.

[0076] The early warning module makes early warning judgments on the wind turbine operation data according to the early warning threshold, and then issues early warnings. Specifically, the following operations are performed:

[0077] The abnormal data ratio of the wind turbine operating data is compared with the warning threshold. If the abnormal data ratio exceeds the warning threshold, the warning is triggered and the system prompts that the unit component is faulty; if the abnormal data ratio does not exceed the warning threshold, the warning is not triggered.

[0078] According to the operating data of the wind turbines that triggered the early warning, the duration of each early warning triggering is analyzed, and the severity of the wind turbine early warning is judged based on the duration; if the duration is less than one hour, it is judged as a level one warning; if the duration is greater than one hour and less than two hours, it is judged as a level two warning; if the duration is greater than two hours, it is judged as a level three warning.

[0079] The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are covered by the protection scope of the present invention.

Claims

1. A fault warning modeling method for a wind turbine generator set, characterized in that: The following steps are involved: S1. Collect wind turbine operation data, pre-process the collected wind turbine operation data, and divide it into a training sample set and a verification sample set. Specifically, perform the following operations: The operation data of the wind turbines in normal operation and abnormal operation with a span of one month are obtained respectively, with a sampling period of 1s; the collected wind turbine data are preprocessed, and the first 80% of the collected wind turbine data are used as the training sample set, and the remaining 20% ​​of the data are used as the verification sample set; S2, input the training sample set into the reliefF algorithm, obtain the feature influence value of each label point data, use the feature influence value as the weight of each label point data, and filter out each label point data according to the weight; S3. Input the filtered label point data into the SVM model for training and build an early warning model. Specifically, perform the following operations: Input the filtered label point data into the SVM model, train the SVM model, and the trained SVM model is the early warning model, which is saved; S4, inputting the operating data of multiple groups of faulty wind turbines into the early warning model, obtaining the abnormal data proportions corresponding to the multiple groups of faulty wind turbines, and calculating the average value of all abnormal data proportions to obtain the early warning threshold, including the following steps: S401, after randomly taking the operation data of the faulty wind turbine set with a data length of 1 day and inputting it into the early warning model, the proportion of abnormal data judged as abnormal data in the operation data of the wind turbine set corresponding to that day is obtained; S402, repeating step S401 for multiple times, and taking the average value of all the abnormal data proportions obtained as the warning threshold; S5. Input the wind turbine operation data currently required to be detected into the early warning model, compare the abnormal data ratio of the wind turbine with the early warning threshold, and determine whether the component of the wind turbine is faulty. Specifically, perform the following operations: After the wind turbine operating data currently required to be detected is input into the early warning model, the abnormal data ratio of the wind turbine operating data is obtained. If the abnormal data ratio exceeds the early warning threshold, an early warning is triggered and the system prompts that the unit component is faulty; if the abnormal data ratio does not exceed the early warning threshold, no early warning is triggered; S6. Analyze the duration of each warning triggering according to the wind turbine operation data that triggers the warning to determine the severity of the wind turbine warning; S7. Input the verification sample set into the early warning model to test the effectiveness of the early warning model.

2. A fault warning modeling method for a wind turbine according to claim 1, characterized in that: In step S6, the following operations are specifically performed: According to the wind turbine operation data that triggered the warning, analyze the duration of each warning trigger, and judge the severity of the wind turbine warning based on the duration; if the duration is less than one hour, it is judged as a first-level warning; if the duration is greater than one hour and less than two hours, it is judged as a second-level warning; If it lasts longer than two hours, it will be considered a Level 3 warning.

3. A fault warning modeling system for wind turbines, characterized in that: include: The operation data acquisition module acquires the operation data of the wind turbine in normal operation and abnormal operation with a span of one month, with a sampling period of 1s; Preprocess the collected wind turbine data, use the first 80% of the collected wind turbine data as the training sample set, and the remaining 20% ​​of the data as the verification sample set; The modeling label acquisition module inputs the training sample set into the reliefF algorithm to obtain the feature influence value of each label point data, uses the feature influence value as the weight of each label point data, and selects each label point data according to the weight; The model training module inputs the filtered corresponding label point data as a training set into the SVM model to build an early warning model. Specifically, the following operations are performed: Input the filtered label point data into the SVM model, train the SVM model, and the trained SVM model is the early warning model, which is saved; The threshold determination module inputs multiple groups of faulty wind turbine operation data into the early warning model, obtains the abnormal data proportions corresponding to the multiple groups of faulty wind turbines, and calculates the average value of all abnormal data proportions to obtain the early warning threshold, including the following steps: S401, after randomly taking the operation data of the faulty wind turbine set with a data length of 1 day and inputting it into the early warning model, the proportion of abnormal data judged as abnormal data in the operation data of the wind turbine set corresponding to that day is obtained; S402, repeating step S401 for multiple times, and taking the average value of all the abnormal data proportions obtained as the warning threshold; The early warning module inputs the wind turbine operation data currently required to be detected into the early warning model, compares the abnormal data ratio of the wind turbine with the early warning threshold, determines whether the components of the wind turbine are faulty, and analyzes the duration of each early warning trigger according to the wind turbine operation data that triggers the early warning to determine the severity of the wind turbine early warning. Specifically, the following operations are performed: After inputting the wind turbine operating data that needs to be detected currently into the early warning model, the abnormal data ratio of the wind turbine operating data is obtained. If the abnormal data ratio exceeds the early warning threshold, an early warning is triggered and the system prompts a unit component failure; if the abnormal data ratio does not exceed the early warning threshold, no early warning is triggered.

4. A fault warning modeling system for wind turbines according to claim 3, characterized in that: The early warning module specifically performs the following operations: According to the wind turbine operation data that triggered the warning, analyze the duration of each warning trigger, and judge the severity of the wind turbine warning based on the duration; if the duration is less than one hour, it is judged as a first-level warning; if the duration is greater than one hour and less than two hours, it is judged as a second-level warning; If it lasts longer than two hours, it will be considered a Level 3 warning.

Images