A multi-wind-turbine operation characteristic cooperative early warning method based on partition clustering

Abstract

The application relates to a kind of multi-wind turbine operation characteristic cooperative early warning methods based on partition clustering, through the multi-wind turbine operation characteristic cooperative early warning method based on partition clustering, first, the obtained operation parameter is data cleaning;Second, the characteristic curve is obtained according to the clustering method;Then, the distance matrix is constructed according to the characteristic curve, and the deviation degree of the characteristic curve is calculated;Finally, according to the deviation degree threshold index, the unit operation characteristic is early warned;The method realizes more sensitive and more accurate identification of abnormal units deviating from normal behavior mode at the group level, which is an innovation in the wind turbine fault early warning method.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine generator condition monitoring and fault early warning technology, specifically involving a collaborative early warning method for the operating characteristics of multiple wind turbine generators based on partitioned clustering. Background Technology

[0002] With the large-scale development of the wind power industry, wind farms typically consist of dozens or even hundreds of wind turbine units. To ensure the safe, efficient, and stable operation of these units, early warning systems have become indispensable. Currently, the monitoring of wind turbine operating status and fault early warning mainly focuses on the individual unit level, that is, determining its health status by analyzing data from the SCADA (Supervisory Control and Data Acquisition) system of a single unit, such as vibration, temperature, and power. This analysis method, which only focuses on individual unit data, cannot capture such group characteristics. If a uniform threshold is used, it will generate a large number of false alarms for normal units that are inherently disadvantaged. Moreover, some slowly developing early faults show minimal changes in individual unit data, easily masked by noise, and difficult for individual unit early warning models to detect in a timely manner. While some existing research attempts to analyze wind farms as a whole, the grouping strategies are rigid and cannot adapt to the dynamically changing operating conditions of wind farms; the accuracy and timeliness of early warnings still need improvement.

[0003] In other words, existing early warning systems mainly rely on single-unit SCADA data analysis (such as vibration, temperature, power, etc.), which has significant limitations:

[0004] 1. Lack of group characteristics: Single-unit analysis cannot capture the dynamic correlation between units, and is prone to false alarms for normal units in inherently unfavorable positions (such as terrain differences);

[0005] 2. Early faults are often missed: Slow-developing faults show only slight changes in single-machine data and are easily masked by noise.

[0006] 3. Rigid strategy: Existing grouping methods are difficult to adapt to real-time changes in wind farm operating conditions, resulting in insufficient accuracy and timeliness of early warnings.

[0007] Therefore, improvement and innovation are imperative. Summary of the Invention

[0008] In view of the above situation and to overcome the shortcomings of the existing technology, the purpose of this invention is to provide a method that can reflect the dynamic operation correlation of wind farms and realize sensitive and accurate early warning of unit anomalies. It can cluster based on the similarity of the real-time operating characteristics of the units, thereby more sensitively and accurately identifying abnormal units that deviate from the normal behavior pattern at the group level.

[0009] The technical solution provided by this invention is: a collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering, comprising the following steps:

[0010] S1. Obtain multi-dimensional operating parameters of all wind turbines in real time from the wind farm's SCADA system within a predetermined time window to form an initial dataset;

[0011] S2. Perform data cleaning on the initial dataset, including removing downtime data and sensor anomalies, to form a standardized dataset;

[0012] Data cleaning rules: 1. Data will be removed when the unit is in a non-grid-connected state; 2. Data will be removed when the wind speed exceeds 60m / s or is less than 0m / s; 3. Data will be removed when the power exceeds 1.2 times the rated power; 4. Data will be removed when the collected operating parameters do not change within 5 minutes.

[0013] S3. Based on the standardized dataset, and according to different power partitions, such as partitioning at 100kW intervals, with the center point of the last power partition being the rated power, the DBSCAN clustering method is used to extract feature vectors for each unit that can comprehensively characterize its operating status. , represents the feature point corresponding to the nth partition of the mth unit;

[0014] The feature vectors are shown in the table below:

[0015]

[0016] S4. Calculate the distance between the characteristic curve of unit i and the characteristic curve of unit j. :

[0017]

[0018] in, The feature point corresponding to the nth partition of unit j The feature point corresponding to the nth partition of unit i;

[0019] S5. Construct the feature curve distance matrix D:

[0020]

[0021] in, The distance between the characteristic curve of unit i and the characteristic curve of unit j; ,when hour, ;

[0022] S6. Sequentially remove the characteristic curves of unit i to obtain the characteristic curve sub-distance matrix. :

[0023]

[0024] in, The distance between the characteristic curve of unit i+1 and the characteristic curve of unit j+1;

[0025] S7. Calculate the sub-distance matrix Average distance :

[0026]

[0027] in, It is the sub-distance matrix of the characteristic curve. The value in the k-th row and l-th column; This represents the total number of generating units.

[0028] S8. Calculate the sub-distance matrix Distance from standard deviation :

[0029]

[0030] in, Sub-distance matrix Average distance It is the sub-distance matrix of the characteristic curve. The value in the k-th row and l-th column; This represents the total number of generating units.

[0031] S9. Set the threshold coefficient ;

[0032] S10, Calculate the sub-distance matrix Concentration :

[0033]

[0034] in, Sub-distance matrix Average distance Sub-distance matrix Distance from standard deviation This is the threshold coefficient;

[0035] S11, Sub-distance matrix Concentration normalization :

[0036]

[0037] in, Sub-distance matrix Concentration Sub-distance matrix Concentration;

[0038] S12, the deviation of the characteristic curve of unit i is set to... :

[0039]

[0040] in, Sub-distance matrix Concentration normalization;

[0041] S13. Set a deviation threshold index based on the deviation of the unit characteristic curve. If the deviation of the unit characteristic curve exceeds the threshold index, an early warning will be issued.

[0042] Preferably, the operating parameters include: unit status, wind speed, active power, generator speed, impeller speed, pitch angle, ambient temperature, gearbox oil sump temperature, main shaft temperature, generator winding temperature, generator bearing temperature, and pitch motor temperature.

[0043] The unit status includes grid-connected status and off-grid status, where grid-connected status is the wind turbine generating power and off-grid status is the unit not generating power.

[0044] Preferably, the threshold coefficient in step S9 The value range is (0.5, 3), and the principle for determining the value is that the deviation of the characteristic curve can effectively distinguish abnormal units.

[0045] Preferably, the deviation threshold index in step S13 is 1.5-2 times the average deviation of the unit characteristic curve obtained in step S12.

[0046] This invention employs a collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering. First, the acquired operating parameters are cleaned. Second, feature curves are obtained using clustering methods. Then, a distance matrix is ​​constructed based on the feature curves, and the deviation of the feature curves is calculated. Finally, an early warning is issued for the unit operating characteristics based on a deviation threshold index. This method achieves more sensitive and accurate identification of abnormal units deviating from normal behavior patterns at the group level, representing an innovation in wind turbine fault early warning methods. Compared with existing technologies, it adaptively divides unit clusters based on real-time operating characteristic similarity (non-fixed grouping), overcoming the challenges of dynamic operating conditions. The impact of state changes is mitigated; by constructing a distance matrix and calculating the concentration of sub-matrixes, the deviation of individual units is quantified at the group level, improving the ability to capture weak abnormal signals; DBSCAN clustering feature points are extracted according to power ranges to accurately characterize the correlation of the operating state of the units under different loads; group collaborative analysis can identify slow faults that are difficult to detect by single-unit models, improving the sensitivity of early faults; dynamic clustering avoids false alarms for normal units with inherent differences in operating characteristics, reducing the false alarm rate; the partition feature extraction and distance matrix update mechanism responds to changes in wind farm operating conditions in real time, with strong adaptability, which is an innovation in wind turbine fault early warning methods. Attached Figure Description

[0047] Figure 1 is a schematic diagram of the collaborative early warning process for the operating characteristics of multiple wind turbine units based on partitioned clustering according to the present invention.

[0048] Figure 2 is a feature curve diagram of an embodiment of the present invention.

[0049] Figure 3 is a feature curve deviation diagram of an embodiment of the present invention. Detailed Implementation

[0050] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0051] See Figure 1 This is a schematic diagram of a collaborative early warning process for the operating characteristics of multiple wind turbine units based on partitioned clustering in this embodiment, including the following steps:

[0052] S1. Taking a wind farm in Henan Province as an example, this wind farm has five 3MW wind turbine units. The operating parameters of all wind turbine units over the past 15 days are obtained in real-time from the wind farm's SCADA system to form an initial dataset. These operating parameters include: wind speed, active power, generator speed, pitch angle, and gearbox oil sump temperature.

[0053] S2. Perform data cleaning on the initial dataset, including removing shutdown data and sensor anomalies, to form a standardized dataset. Data cleaning rules: 1. Data when the unit is in a non-grid-connected state is removed; 2. Data when the wind speed exceeds 60m / s or is less than 0m / s is removed; 3. Data when the power exceeds 1.2 times the rated power is removed; 4. Data when the collected operating parameters do not change within 5 minutes is removed.

[0054] S3. Based on the standardized dataset, partition the dataset according to power, with a partition interval of 100kW. The power partitions are as follows:

[0055] The DBSCAN clustering method is used to extract feature vectors for each unit that comprehensively characterize its operating status. (See also...) Figure 2 ;

[0056] S4. Calculate the distance between the characteristic curves of the two units;

[0057] S5. Construct the feature curve distance matrix D:

[0058]

[0059] S6. Remove the characteristic curves of the generator set one by one to obtain the characteristic curve sub-distance matrix;

[0060] S7. Calculate the average distance of the sub-distance matrices sequentially;

[0061] S8. Calculate the standard deviation of the distance in each sub-distance matrix in turn;

[0062] S9. Set the threshold coefficient alpha to 2;

[0063] S10. Calculate the concentration of the sub-distance matrix sequentially:

[0064]

[0065] S11, Concentration normalization of the sub-distance matrix:

[0066]

[0067] S12, Deviation of Unit Characteristic Curve (participant) Figure 3 The deviation of the characteristic curve of Unit 1 is the reciprocal of the sub-distance matrix after centralized normalization, which is 1 / 1.2612 = 0.7929. Using the same method, the deviations of the characteristic curves of Units 2, 3, 4, and 5 are 0.8621, 0.8752, 0.8518, and 3.8112, respectively.

[0068]

[0069] S13. Set the deviation threshold index to 1.5. The deviation of the characteristic curve of Unit 5 is 3.8112, which exceeds the threshold index. Unit 5 reports an abnormal temperature warning for the gearbox oil sump.

[0070] S16. Upon on-site inspection, it was found that the surface of the heat sink of the gearbox cooler of Unit 5 was severely dusty, resulting in poor ventilation and thus causing the gearbox oil sump temperature to be higher than that of other units. After cleaning, the temperature returned to normal.

Claims

1. A collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering, characterized in that, Includes the following steps: S1. Obtain multi-dimensional operating parameters of all wind turbines in real time from the wind farm's SCADA system within a predetermined time window to form an initial dataset; S2. Perform data cleaning on the initial dataset, including removing downtime data and sensor anomalies, to form a standardized dataset; S3. Based on the standardized dataset and according to different power zones, extract feature vectors that comprehensively characterize the operating status of each unit using the DBSCAN clustering method. , represents the feature point corresponding to the nth partition of the mth unit; S4. Calculate the distance between the characteristic curve of unit i and the characteristic curve of unit j. : in, The feature point corresponding to the nth partition of unit j The feature point corresponding to the nth partition of unit i; S5. Construct the feature curve distance matrix D: in, The distance between the characteristic curve of unit i and the characteristic curve of unit j; ,when hour, ; S6. Sequentially remove the characteristic curves of unit i to obtain the characteristic curve sub-distance matrix. : in, The distance between the characteristic curve of unit i+1 and the characteristic curve of unit j+1; S7. Calculate the sub-distance matrix Average distance : in, It is the sub-distance matrix of the characteristic curve. The value in the k-th row and l-th column; This represents the total number of generating units. S8. Calculate the sub-distance matrix Distance from standard deviation : in, Sub-distance matrix Average distance It is the sub-distance matrix of the characteristic curve. The value in the k-th row and l-th column; This represents the total number of generating units. S9. Set the threshold coefficient ; S10, Calculate the sub-distance matrix Concentration : in, Sub-distance matrix Average distance Sub-distance matrix Distance from standard deviation This is the threshold coefficient; S11, Sub-distance matrix Concentration normalization : in, Sub-distance matrix Concentration Sub-distance matrix Concentration; S12, the deviation of the characteristic curve of unit i is set to... : in, Sub-distance matrix Concentration normalization; S13. Set a deviation threshold index based on the deviation of the unit characteristic curve. If the deviation of the unit characteristic curve exceeds the threshold index, an early warning will be issued.

2. The collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering according to claim 1, characterized in that, The operating parameters include: unit status, wind speed, active power, generator speed, impeller speed, pitch angle, ambient temperature, gearbox oil sump temperature, main shaft temperature, generator winding temperature, generator bearing temperature, and pitch motor temperature.

3. The collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering according to claim 1, characterized in that, The threshold coefficient in step S9 The value range is (0.5, 3).

4. The collaborative early warning method for the operating characteristics of multiple wind turbine units based on partitioned clustering according to claim 1, characterized in that, The deviation threshold index in step S13 is 1.5-2 times the average deviation of the unit characteristic curve obtained in step S12.

Images