Transformer winding thermal defect detection method based on Hypergrid-XGBoost

Abstract

The invention belongs to the technical field of transformer winding defect detection, and particularly relates to a transformer winding thermal defect detection method based on Hypergrid-XGBoost. Comprising the following steps: S1, building a simulation model of the transformer and a winding on a Comsol platform according to physical parameters of a target oil immersed transformer, and verifying the effectiveness of the simulation model through a temperature rise test; s2, arranging optical fiber sensor nodes, performing abnormal value test on data transmitted by a sensor by using a Hypergrid strategy, and constructing a winding actual measurement temperature rise matrix; and S3, establishing a temperature rise matrix of the winding in a normal state, and determining the position of the thermal defect of the winding. S4, establishing an overheat point temperature rise model by using an XGBoost algorithm based on the temperature data, and determining the severity of the winding thermal defect; according to the invention, Hypergrid abnormal value detection and an XGBoost model learning data processing technology are applied to an actual scene of winding thermal defect detection, high-precision positioning of a transformer overheat point is realized, the severity of temperature rise of the overheat point is identified, and a direct basis is provided for formulation and implementation of transformer protection measures.

Description

Technical Field

[0001] The present invention belongs to the technical field of transformer winding defect detection, and in particular relates to a transformer winding thermal defect detection method based on Hypergrid-XGBoost. Background Art

[0002] Winding is one of the most important and most vulnerable internal components of oil-immersed transformers. Insulation aging and short circuits between turns or layers of windings are the main factors causing the temperature rise of windings. The increase in local winding temperature caused by winding oil channel blockage or winding short circuit is called winding thermal defect. If these thermal defects are not handled in time, the local winding temperature is too high, which may cause the transformer to explode or catch fire, endangering the stability of power grid operation and the safety of life and property. Therefore, it is necessary to identify the location and severity of thermal defects of transformer windings in real time, discover potential faults of transformer windings in time, and take corresponding measures to eliminate them to ensure the safe operation of transformers.

[0003] The commonly used diagnostic methods for thermal defects of oil-immersed transformers and their windings are mainly dissolved gas analysis (DGA), infrared temperature measurement, magnetic leakage, frequency domain response, current waveform, and data-mechanism drive. It is difficult to identify the specific location of thermal defects by the dissolved gas analysis method in oil, and infrared temperature measurement only measures the surface temperature of the equipment, which is difficult to accurately reflect the location and severity of thermal defects inside the equipment. Magnetic leakage, frequency domain response and current waveform are easily affected by the external environment and operating conditions. In the data-mechanism drive method, the sensor will inevitably collect abnormal values, and the spatiotemporal validity of the sensor data needs to be detected, which will greatly improve the accuracy of winding defect detection. In view of this, the present invention applies data processing technology to transformer winding fault diagnosis, and proposes a transformer winding thermal defect detection method based on Hypergrid-XGBoost. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and to provide a transformer winding thermal defect detection method based on Hypergrid-XGBoost to locate the location of the thermal defect in the winding and determine the severity of the thermal defect.

[0005] The object of the present invention is achieved by: a transformer winding thermal defect detection method based on Hypergrid-XGBoost, which comprises the following steps:

[0006] S1. Build a simulation model of the transformer and windings on the Comsol platform based on the physical parameters of the target oil-immersed transformer and verify the effectiveness of the simulation model through a temperature rise test;

[0007] S2, partition the transformer winding, arrange the optical fiber sensor nodes, use the Hypergrid strategy to test the outliers of the data transmitted by the sensor, use the MF method to estimate the missing values, and construct the measured temperature rise matrix of the winding;

[0008] S3. On the transformer and winding simulation platform, a temperature rise matrix of the winding under normal conditions is established through data quadratic fitting. The location of the thermal defect of the winding is determined based on the measured temperature rise matrix of the winding and the temperature rise matrix of the winding under normal conditions.

[0009] S4. Based on the temperature data, the XGBoost algorithm is used to establish a hot spot temperature rise model to determine the severity of the thermal defects of the winding.

[0010] In step S2, performing an outlier check on the data transmitted by the sensor using the Hypergrid strategy includes the following steps:

[0011] S21, constructing information vector T according to the temperature rise data collected by the temperature sensor mn , and the information vector T mn standardization;

[0012] S22, the standardized information vector T mn Mapping to a hypercube Computing Hypercubes and the amount of data in its neighborhood, identifying outliers based on standard documentation;

[0013] S23, each zone of the winding is defined as a data cluster, and the standard file is updated by using intermittent learning in each zone of the winding, and the entire winding executes a continuous learning strategy to update the entire standard file;

[0014] S24. All partitions of the winding perform real-time outlier detection and remove outliers.

[0015] In step S3, establishing a temperature rise matrix of the winding under normal state by quadratic fitting of data on the transformer and winding simulation platform includes the following steps:

[0016] S31, defining the layers of the windings, and simulating the temperature rise of the temperature measuring points using a simulation model;

[0017] S32, under the unchanged working condition, simulate and obtain the temperature rise simulation value of the temperature test point at the current time k and the previous time k-1 under the normal state of the winding;

[0018] S33, according to the temperature rise information vector T mn The definition of the temperature rise information vector T of the winding under normal conditions is calculated mn The simulation value of

[0019] S34, perform quadratic fitting on the load rate θ and the oil flow velocity v at the winding inlet to obtain the winding normal state temperature rise information vector T mn The fitted value of

[0020] S35. Interpolate the simulation data and the fitting data to obtain the temperature rise value of the test point in the normal state.

[0021] The step S4 of establishing the hotspot temperature rise model using the XGBoost algorithm includes the following steps:

[0022] S41, determining the number of boosted trees and the fault tolerance value, as well as the area where the hotspot is located, i.e., the target detection area;

[0023] S42, determining a value set of a load factor and an oil flow velocity at a winding inlet;

[0024] S43, setting a hot spot in the target detection area, and simulating the temperature rise value of the temperature measuring point at different temperatures of the hot spot under different working conditions;

[0025] S44, using the simulation data obtained in step S43 as a training set, applying the XGBoost learning algorithm, and constructing a hot spot temperature rise value mapping with the load rate, the oil flow velocity at the winding inlet, and the temperature rise at the temperature measuring point as independent variables.

[0026] Beneficial effects of the present invention: The transformer winding thermal defect detection method based on Hypergrid-XGBoost of the present invention has two functions of determining the location of the hot spot and evaluating the severity of the temperature rise of the hot spot. First, a simulation model of the transformer and its winding is established according to the physical parameters of the physical transformer, and the effectiveness of the simulation model is verified by a temperature rise test. Then, the thermal defect detection method is initialized, the threshold and the local standard matrix are given, the winding is partitioned, and the temperature test point closest to the oil flow inlet of each partition is designated as the cluster head, and the sensor is deployed in the physical winding. Next, the location of the thermal defect is determined. The principle of determining the location of the thermal defect is very simple. With the help of the transformer simulation model and quadratic fitting, the normal state temperature rise matrix of each partition of the winding is obtained, and the temperature rise matrix of the winding is constructed using sensor data. The two matrices are subtracted and compared with the threshold matrix to determine the partition where the hot spot is located. In practice, sensor data has abnormal values. In order to improve the accuracy of identifying the location of overheating spots and the severity of temperature rise, the temperature rise feature vector of the winding is constructed with the help of sensor data. The feature vector is standardized by using mean-variance to prevent the grid from being too dense. The feature vector is mapped to Hypergrid, and the amount of data in Hypergrid and its L1 and L2 neighborhoods is calculated and compared with the standard file to detect data abnormal values. Because the standard file is not well known, the sensor and cluster head need to perform intermittent learning to update the standard file. The various regions of the winding are continuously learned, and the MF estimation is performed on the missing data values. The data is continuously output to the computer to form the measured temperature rise matrix. Finally, the severity of the temperature rise of the overheating spot is identified. The data of load rate, inlet oil flow velocity, overheating spot and temperature measurement point obtained by the simulation model are used as the training set to learn the XGBoost model of the overheating spot. The measured load rate, inlet oil flow velocity and temperature measurement point data are substituted into the XGBoost model to estimate the value of the overheating spot temperature rise and judge the severity of overheating at the fault point.

[0027] The present invention applies Hypergrid outlier detection and XGBoost model learning data processing technology to the actual scenario of winding thermal defect detection, realizes high-precision positioning of transformer hot spots, and identifies the severity of hot spot temperature rise, providing a direct basis for the formulation and implementation of transformer protection measures. A feature vector characterizing the variation characteristics of winding temperature rise is constructed, achieving a more comprehensive and detailed description of winding temperature rise. The Hypergrid method is used to estimate the outliers of the feature vector data, which improves the accuracy of the diagnosis of the location of winding thermal defects; the XGBoost model of the winding hot spot is trained according to the data set, and the method of inputting the measured data into the model to obtain the temperature rise of the hot spot is simple, easy to operate and has high accuracy, and can be deployed in a real transformer fault detection system. The effectiveness of the detection method has been verified through experiments in Comsol simulation software and transformer entities. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0029] Figure 1 Flowchart for determining hot spot locations in thermal defect detection of transformer windings based on Hypergrid-XGBoost.

[0030] Figure 2 Flowchart for estimating overheating degree in transformer winding thermal defect detection based on Hypergrid-XGBoost.

[0031] Figure 3 It is a simulation model of SSZ11-25000 / 110 transformer and windings.

[0032] Figure 4 This is the flow chart of Hypergrid data outlier detection.

[0033] Figure 5 Flowchart for estimating missing values ​​using the matrix factorization (MF) method.

[0034] Figure 6 Schematic diagram of the prediction error of the severity of thermal defects in windings. DETAILED DESCRIPTION

[0035] The present invention will be further described below in conjunction with the accompanying drawings.

[0036] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The transformer winding thermal defect detection method based on Hypergrid-XGBoost of the present invention comprises the following steps:

[0039] S1. Build a simulation model of the transformer and windings on the Comsol platform based on the physical parameters of the target oil-immersed transformer and verify the effectiveness of the simulation model through a temperature rise test;

[0040] S2, partition the transformer winding, arrange the optical fiber sensor nodes, use the Hypergrid strategy to test the outliers of the data transmitted by the sensor, use the MF method to estimate the missing values, and construct the measured temperature rise matrix of the winding;

[0041] S3. On the transformer and winding simulation platform, a temperature rise matrix of the winding under normal conditions is established through data quadratic fitting. The location of the thermal defect of the winding is determined based on the measured temperature rise matrix of the winding and the temperature rise matrix of the winding under normal conditions.

[0042] S4. Based on the temperature data, the XGBoost algorithm is used to establish a hot spot temperature rise model to determine the severity of the thermal defects of the winding.

[0043] Furthermore, in step S2, performing an outlier check on the data transmitted by the sensor using the Hypergrid strategy includes the following steps:

[0044] S21, constructing information vector T according to the temperature rise data collected by the temperature sensor mn , and the information vector T mn standardization;

[0045] S22, the standardized information vector T mn Mapping to a hypercube Computing Hypercubes and the amount of data in its neighborhood, identifying outliers based on standard documentation;

[0046] S23, each zone of the winding is defined as a data cluster, and the standard file is updated by using intermittent learning in each zone of the winding, and the entire winding executes a continuous learning strategy to update the entire standard file;

[0047] S24. All partitions of the winding perform real-time outlier detection and remove outliers.

[0048] Furthermore, in step S3, establishing a temperature rise matrix of the winding under normal state by quadratic fitting of data on the transformer and winding simulation platform includes the following steps:

[0049] S31, defining the layers of the windings, and simulating the temperature rise of the temperature measuring points using a simulation model;

[0050] S32, under the unchanged working condition, simulate and obtain the temperature rise simulation value of the temperature test point at the current time k and the previous time k-1 under the normal state of the winding;

[0051] S33, according to the temperature rise information vector T mnThe definition of the temperature rise information vector T of the winding under normal conditions is calculated mn The simulation value of

[0052] S34, perform quadratic fitting on the load rate θ and the oil flow velocity v at the winding inlet to obtain the winding normal state temperature rise information vector T mn The fitted value of

[0053] S35. Interpolate the simulation data and the fitting data to obtain the temperature rise value of the test point in the normal state.

[0054] Furthermore, in step S4, establishing the hot spot temperature rise model using the XGBoost algorithm includes the following steps:

[0055] S41, determining the number of boosted trees and the fault tolerance value, as well as the area where the hotspot is located, i.e., the target detection area;

[0056] S42, determining a value set of a load factor and an oil flow velocity at a winding inlet;

[0057] S43, setting a hot spot in the target detection area, and simulating the temperature rise value of the temperature measuring point at different temperatures of the hot spot under different working conditions;

[0058] S44, using the simulation data obtained in step S43 as a training set, applying the XGBoost learning algorithm, and constructing a hot spot temperature rise value mapping with the load rate, the oil flow velocity at the winding inlet, and the temperature rise at the temperature measuring point as independent variables.

[0059] The present invention proposes a detection method for thermal defects of transformer windings based on Hypergrid-XGBoost, which can locate the location of thermal defects in the windings and determine the severity of the thermal defects. A simulation model of the transformer and its windings is established and the validity of the model is verified. The transformer windings are partitioned, and optical fiber sensor nodes are arranged. The Hypergrid strategy is used to implement outlier detection on the spatiotemporal data transmitted by the sensors, and the MF method is used to estimate the missing values. The measured temperature rise matrix of the windings is constructed. The temperature rise matrix of the windings in a normal state is established by quadratic fitting of the data on the transformer and winding simulation platform. The location of the thermal defects of the windings is determined according to the difference matrix between the measured temperature rise matrix of the windings and the temperature rise matrix of the windings in a normal state. The temperature rise model of the overheat point is learned by the XGBoost algorithm, and the severity of the thermal defects of the windings is determined according to the predicted value of the temperature rise model.

[0060] The main idea of ​​Hypergrid outlier detection is to map the obtained spatiotemporal data to a hypercube. The dimension of the data determines the number of side lengths of the hypercube. In fact, the data is mapped to a finite-dimensional space. Next, a neighborhood is defined for the points in the space, and a threshold is defined based on experience. If the number of points in the neighborhood of the tested point exceeds the threshold, the tested point is considered normal, otherwise, the tested point is considered abnormal.

[0061] The idea of ​​XGBoost is to integrate many weak classifiers together to form a strong classifier. Its main element is the decision tree, which is a weak classifier. There is a sequence between the decision trees: the generation of the latter decision tree will take into account the prediction result of the previous decision tree, that is, the deviation of the previous decision tree is taken into account, so that the training samples that the previous decision tree made mistakes will receive more attention in the future, and then the next decision tree will be trained based on the adjusted sample distribution. These decision trees constitute XGBoost.

[0062] The present invention is described in detail below by taking a specific type of transformer as an example.

[0063] (1) Take SSZ11-25000 / 110 transformer as an example. Figure 2 As shown, its physical parameters are shown in Table 1. The implementation process of the detection method of winding thermal defects is explained through the A-phase low-voltage winding of the actual transformer. The total loss of the winding, that is, the sum of the total eddy current loss and the resistance loss, is used as the heat source of the winding. The A-phase low-voltage winding has 5 oil baffles, and the winding is divided into 4 partitions by the oil baffles. A temperature analysis point is set in the vertical oil channel on the outlet side of each partition, and 3 hot spots are set, which are respectively recorded as the subscript represents the phase, the first digit of the subscript represents the low voltage, and the second digit represents the analysis point serial number. At this time, the winding has a total of 13 temperature analysis points, and optical fiber temperature sensors are embedded in these 13 temperature analysis points. The area between two adjacent temperature analysis points is defined as one layer, with a total of 12 layers. In actual application scenarios, staff can customize the number of temperature analysis points according to detection requirements.

[0064] Table 1 Part of the physical parameters of the transformer

[0065]

[0066] (2) A simulation model of the physical transformer and its windings in step 1 is established on the Comsol software platform. Sensors are arranged in the simulation model according to the distribution of sensors in the physical transformer. The validity of the simulation model is verified by a temperature rise test.

[0067] (3) Perform outlier inspection on the spatiotemporal data stream transmitted by the sensors arranged in the physical transformer.

[0068] In practical applications, sensors will inevitably collect abnormal data. It is necessary to check the data collected by the temperature sensor in the physical transformer and remove the abnormal data to improve the accuracy of the collected data. The present invention uses a hypergrid structure to check the data collected by the optical fiber sensor. Figure 4 The sensor is at the temperature analysis point M mn The temperature rise value collected at time k is recorded as Define a new collection information vector T mn :

[0069]

[0070] in T mn Mapping to a hypercube where u1u2…u q is the coordinate index of the hypercube CC, represented by T mn The dimension of the hypercube is a five-dimensional space. Assume that the side length and diagonal length of the cube are h and d respectively. Since the dimension of the data feature space is 5, The mapping relationship is as follows:

[0071] i=1,2,…5,C is a constant, and C>|min(T mn1 )|

[0072] In order to avoid the hypercube grid being too dense and low-intensity noise not being detected, and to improve the detection accuracy, the data needs to be standardized. Take the data standardization process as an example. First, calculate the information vector T mn The mean of the elements in and variance Secondly, T mn Each element in constructs a new variable, The standard form of the information vector is obtained. For simplicity and without loss of generality, the standardized information vector T mn Continue to express as

[0073]

[0074] Define the neighborhood of the detected point. The present invention adopts the following neighborhood form, denoted as

[0075] After mapping the sensor spatiotemporal data to the 5-dimensional hypercube space, search for data points in a given neighborhood and compare the number of data points with the threshold. If the value is greater than the threshold, the measured data point is a normal value, otherwise, the measured data point is judged as an abnormal value. The following describes how to determine the size of the threshold and the range of the neighborhood.

[0076] (3.1) First, define each zone of the winding as a network cluster. The A-phase low-voltage winding has 4 sensor network clusters. The temperature measurement point in each zone of the winding is recorded as a sensor node (SN, also called a cluster node), and the temperature measurement point at the top of each zone closest to the oil flow outlet is the cluster head node (CH). Then, each network cluster of the winding has 9 cluster nodes and 1 cluster head. The winding has 4 cluster heads in total. The convergence point of the 4 cluster head nodes is recorded as the convergence node, and the convergence node transmits the data to the computer.

[0077] (3.2) Initialization: At the initial moment, the data of the nodes in the cluster and the cluster head node are directly stored in the local standard file. The data in the default standard file are all normal data. The number of data points in the neighborhood of each data point in the standard file is searched according to the neighborhood form, and the minimum value is recorded as the threshold.

[0078] (3.3) When the amount of data in the locally stored standard file exceeds the predefined value, the cluster nodes and cluster head nodes start the distributed learning mechanism. That is, a time window w is defined in advance. After the first learning, the cluster nodes and cluster head nodes will learn again after a period of time w. This kind of learning belongs to intermittent learning. During each learning process, the local standard file is updated with the data stream collected in the time window, and the new threshold and neighborhood form are determined according to the updated standard file.

[0079] After the nodes in the cluster and the cluster head perform intermittent learning, they synchronously transmit the updated thresholds, neighborhood forms and abnormal data to the sink node, which transmits them to the computer in real time.

[0080] (4) Estimation of missing values ​​of data transmitted by sensors arranged in physical transformers

[0081] Missing data may occur in the sensor data stream. Even after removing the abnormal values ​​in the detected sensor data, missing data may still occur. The present invention estimates the missing data using interpolation methods and fills them in the data stream. The present invention uses matrix factorization (MF) to estimate the missing values. The implementation process is shown in Figure 5 .

[0082] (5) Obtain the measured temperature rise matrix

[0083] The four cluster heads of the transformer A phase low voltage winding aggregate the data estimated by missing values ​​after passing the abnormality test to the computer, and obtain the measured temperature rise matrix of the winding at time k. The matrix is ​​a 5×10 matrix, as follows:

[0084]

[0085] (6) Calculate the temperature rise matrix of the winding under normal conditions on the simulation platform

[0086] A point between two temperature test points is randomly selected as the hot spot and recorded as G. The part between the two temperature test points is called a layer. Different temperatures are set for the hot spots in the layer, and the temperature rise of the winding temperature test point under each hot spot temperature value is obtained by simulation. These temperature rises are written in the form of a matrix:

[0087]

[0088] Where T n1i 、T n2i 、T n3i 、T n4i 、T n5i They respectively represent the temperature rise of the simulated test point, the change in temperature rise, and the rate of change of temperature rise between the current moment and the previous moment, i=1,2,…10.

[0089] Considering that the winding temperature rise is affected by the transformer working conditions, the temperature rise at the measuring point is fitted by the load rate θ and the oil flow velocity v at the winding inlet, using the quadratic fitting formula

[0090] T m1i =ω1+ω2v+ω3v 2 +ω4θ+ω5θ 2

[0091] The temperature rise of the test point without thermal defects is the interpolation of the simulation data and the fitting data.

[0092] T i =μT n1i +(1-μ)T m1i ,μ∈(0,1)

[0093] Let T n1i =T 1i Bring it into T ws Get the temperature rise matrix of the winding under normal condition

[0094]

[0095] (7) Determine the location of thermal defects

[0096] Given a threshold matrix,

[0097]

[0098] The measured temperature rise matrix is ​​subtracted from the simulated normal state temperature rise matrix and compared with the threshold matrix. If it is greater than the threshold, a defect exists and the transformer issues an alarm to determine the area where the thermal defect occurs. If it is less than or equal to the threshold, the defect does not exist and the sensor node continues to transmit data at the next moment, repeating the above (3)-(5).

[0099] (8) XGBoost builds hot spot temperature rise model

[0100] The area where the hot spot is located is the target detection area. As mentioned above, 8 temperature test points are deployed in the target detection area. The point of interest is the hot spot in the area. The load rates are 0.2, 0.5, 0.6, 0.8, 1, and 1.1 respectively. The winding inlet flow rate is 0.015, 0.017, 0.020, 0.022, and 0.025 m / s respectively. The hot spot temperatures are set to 200, 260, 400, 630, 700, 810, and 900 respectively. The temperature rise data set of the temperature test point in the target detection area under each working condition at three moments k, k+1, and k+2 is simulated in the simulation model.

[0101] The temperature rise data set of the temperature test point in the target detection area under each working condition at the three moments k, k+1, and k+2 is used as the training set, and the training data set obtained by the simulation model is used to learn the XGBoost model of the temperature rise of the hot spot.

[0102]

[0103] Where M represents the number of boosted trees, f i (x) is a regression tree, which indicates the score of the leaf node. temp indicates the temperature rise of the hot spot. It is a 12-input 1-output mapping.

[0104] (9) Determine the severity of thermal defects

[0105] In the defect matrix, when the transformer issues an overheating alarm, data is collected at the current moment of the alarm to determine the cluster where the overheating point is located based on the location of the overheating point. The sensor data of the cluster where the overheating point is located, the inlet oil flow velocity in the cluster, and the load rate are input into the model in (8) above to calculate the overheating point temperature and give an overheating severity level, based on which protective measures are taken for the transformer.

[0106] The present invention cuts 8 sections of the winding, and the hierarchical partitioning setting of each section is the same as above (1). The transformer winding defect detection method based on Supergrid-XGBoost proposed in the present invention is used to calculate the errors between the positions of 8 hot spots and the severity of overheating. The position error is the layer where the actual position of the hot spot and the judged position are located. The error in thermal severity is the ratio of the difference between the actual overheating value and the predicted overheating value to the actual value. The simulation results show the effectiveness of the algorithm.

[0107] In summary, the transformer winding thermal defect detection method based on Hypergrid-XGBoost of the present invention has two functions of determining the location of the hot spot and evaluating the severity of the temperature rise of the hot spot. First, a simulation model of the transformer and its winding is established according to the physical parameters of the physical transformer, and the effectiveness of the simulation model is verified by a temperature rise test. Then, the thermal defect detection method is initialized, the threshold and the local standard matrix are given, the winding is partitioned, and the temperature test point closest to the oil flow inlet in each partition is designated as the cluster head, and the sensor is deployed in the physical winding. Next, the location of the thermal defect is determined. The principle of determining the location of the thermal defect is very simple. With the help of the transformer simulation model and quadratic fitting, the normal state temperature rise matrix of each partition of the winding is obtained, and the temperature rise matrix of the winding is constructed using sensor data. The two matrices are subtracted and compared with the threshold matrix to determine the partition where the hot spot is located. In practice, sensor data has abnormal values. In order to improve the accuracy of identifying the location of overheating spots and the severity of temperature rise, the temperature rise feature vector of the winding is constructed with the help of sensor data. The feature vector is standardized by using mean-variance to prevent the grid from being too dense. The feature vector is mapped to Hypergrid, and the amount of data in Hypergrid and its L1 and L2 neighborhoods is calculated and compared with the standard file to detect data abnormal values. Because the standard file is not well known, the sensor and cluster head need to perform intermittent learning to update the standard file. The various regions of the winding are continuously learned, and the MF estimation is performed on the missing data values. The data is continuously output to the computer to form the measured temperature rise matrix. Finally, the severity of the temperature rise of the overheating spot is identified. The data of load rate, inlet oil flow velocity, overheating spot and temperature measurement point obtained by the simulation model are used as the training set to learn the XGBoost model of the overheating spot. The measured load rate, inlet oil flow velocity and temperature measurement point data are substituted into the XGBoost model to estimate the value of the overheating spot temperature rise and judge the severity of overheating at the fault point.

[0108] The present invention applies Hypergrid outlier detection and XGBoost model learning data processing technology to the actual scenario of winding thermal defect detection, realizes high-precision positioning of transformer hot spots, and identifies the severity of hot spot temperature rise, providing a direct basis for the formulation and implementation of transformer protection measures. A feature vector characterizing the variation characteristics of winding temperature rise is constructed, achieving a more comprehensive and detailed description of winding temperature rise. The Hypergrid method is used to estimate the outliers of the feature vector data, which improves the accuracy of the diagnosis of the location of winding thermal defects; the XGBoost model of the winding hot spot is trained according to the data set, and the method of inputting the measured data into the model to obtain the temperature rise of the hot spot is simple, easy to operate and has high accuracy, and can be deployed in a real transformer fault detection system. The effectiveness of the detection method has been verified through experiments in Comsol simulation software and transformer entities.

[0109] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0110] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected.

Claims

1. A transformer winding thermal defect detection method based on Hypergrid-XGBoost, characterized in that: The following steps are involved: S1. Build a simulation model of the transformer and windings on the Comsol platform based on the physical parameters of the target oil-immersed transformer and verify the effectiveness of the simulation model through a temperature rise test; S2, partition the transformer winding, arrange the optical fiber sensor nodes, use the Hypergrid strategy to test the outliers of the data transmitted by the sensor, use the MF method to estimate the missing values, and construct the measured temperature rise matrix of the winding; S3. Establish a temperature rise matrix of the winding under normal conditions by quadratic data fitting on the transformer and winding simulation platform, and determine the location of the thermal defect of the winding according to the measured temperature rise matrix of the winding and the temperature rise matrix of the winding under normal conditions; S4. Based on the temperature data, the XGBoost algorithm is used to establish a hot spot temperature rise model to determine the severity of the thermal defects of the winding.

2. The method for detecting thermal defects of transformer windings based on Hypergrid-XGBoost according to claim 1, characterized in that: In step S2, performing an outlier check on the data transmitted by the sensor using the Hypergrid strategy includes the following steps: S21, constructing information vector T according to the temperature rise data collected by the temperature sensor mn , and the information vector T mn standardization; S22, the standardized information vector T mn Mapping to a hypercube Computing Hypercubes and the amount of data in its neighborhood, identifying outliers based on standard documentation; S23, each zone of the winding is defined as a data cluster, and the standard file is updated by using intermittent learning in each zone of the winding, and the entire winding executes a continuous learning strategy to update the entire standard file; S24. All partitions of the winding perform real-time outlier detection and remove outliers.

3. The method for detecting thermal defects of transformer windings based on Hypergrid-XGBoost according to claim 1, characterized in that: In step S3, establishing a temperature rise matrix of the winding under normal state by quadratic fitting of data on the transformer and winding simulation platform includes the following steps: S31, defining the layers of the windings, and simulating the temperature rise of the temperature measuring points using a simulation model; S32, under the unchanged working condition, simulate and obtain the temperature rise simulation value of the temperature test point at the current time k and the previous time k-1 under the normal state of the winding; S33, according to the temperature rise information vector T mn The definition of the temperature rise information vector T of the winding under normal conditions is calculated mn The simulation value of S34, perform quadratic fitting on the load rate θ and the oil flow velocity v at the winding inlet to obtain the winding normal state temperature rise information vector T mn The fitted value of S35. Interpolate the simulation data and the fitting data to obtain the temperature rise value of the test point in the normal state.

4. The method for detecting thermal defects of transformer windings based on Hypergrid-XGBoost according to claim 1, characterized in that: The step S4 of establishing the hotspot temperature rise model using the XGBoost algorithm includes the following steps: S41, determining the number of boosted trees and the fault tolerance value, as well as the area where the hotspot is located, i.e., the target detection area; S42, determining a value set of a load factor and an oil flow velocity at a winding inlet; S43, setting a hot spot in the target detection area, and simulating the temperature rise value of the temperature measuring point at different temperatures of the hot spot under different working conditions; S44, using the simulation data obtained in step S43 as a training set, applying the XGBoost learning algorithm, and constructing a hot spot temperature rise value mapping with the load rate, the oil flow velocity at the winding inlet, and the temperature rise at the temperature measuring point as independent variables.

Images