A dry-type transformer anomaly detection method based on operating condition identification

Abstract

The present invention discloses a dry-type transformer anomaly detection method based on operating condition identification. In the offline model training stage, first, the operating conditions of the dry-type transformer historical monitoring data are identified and divided; secondly, for each operating condition, a self-organizing map network (SOM) network structure and initialization parameters are designed to extract the mapping relationship between the temperatures of the three-phase windings under different operating conditions; then, a random forest (RF) model is trained, and other monitoring quantities are used to predict and model the three-phase winding temperatures; in the anomaly detection stage, if the temperature monitoring data of a phase winding is missing, the missing temperature value is filled using the extracted three-phase winding temperature mapping relationship; further, the predicted temperature of the model at the previous moment is used as the input of the RF model at the next moment to obtain the predicted temperature value of the three-phase winding; finally, the degree of deviation between the actual monitored temperature value of the three-phase winding and the temperature value predicted by the RF model is calculated, and the anomaly detection result is given.

Description

Technical Field

[0001] The present invention belongs to the technical field of power detection, and in particular relates to a dry-type transformer abnormality detection method based on working condition identification. Background Art

[0002] As a core component of power systems, transformers play a vital role in power generation, distribution, transformation, and transmission. Transformers can be categorized as dry-type or oil-immersed based on their insulation type. Due to their compact size, low noise, high operating efficiency, and ease of maintenance, dry-type transformers are widely used in locations with stringent fire and explosion protection requirements, such as high-rise buildings, airports, and train stations. Failures in dry-type transformers are primarily caused by aging of the insulation material. When the temperature of a dry-type transformer winding rises abnormally, the insulation material ages and changes its molecular structure. Even when the temperature returns to normal, the insulation material cannot regain its original structure. Therefore, research on detecting temperature anomalies in dry-type transformer windings has significant engineering value and practical significance.

[0003] The essence of dry-type transformer winding temperature anomaly detection is to detect early-stage thermal failures. This requires establishing an accurate dynamic thermal model of the dry-type transformer to predict the winding temperature under different load conditions. The degree of anomaly is then measured by determining the difference between the measured and predicted winding temperatures. Existing methods for establishing dry-type transformer thermal models primarily include numerical calculation-based methods and artificial intelligence-based methods. Numerical calculation methods can establish relatively accurate dry-type transformer thermal models, but due to the complexity of dry-type transformer structures and the nonlinearity of material parameters, the numerical calculation modeling process is extremely complex and lacks universality.

[0004] The advantage of predicting hotspot temperatures in dry-type transformers based on artificial intelligence algorithms is that they directly establish a relationship between transformer winding temperature and other monitored variables. However, system modeling often fails to account for variations in equipment operating conditions. Dry-type transformer operating conditions typically vary on a daily basis, resulting in different mapping relationships between other monitored variables and winding temperature under different operating conditions. Furthermore, over long-term use, the internal temperature sensors in dry-type transformer windings may miss certain phase winding temperature data during online monitoring due to electromagnetic interference, mechanical compression, and other factors. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing dry-type transformer winding temperature anomaly detection method, the present invention proposes a dry-type transformer anomaly detection method based on operating condition identification. The method can be divided into two stages: model offline training and online anomaly detection. In the model offline training stage, first, the operating conditions of the dry-type transformer historical monitoring data are identified and divided; secondly, for each operating condition, the self-organizing map network (Self-Organizing Maps, SOM) network structure and initialization parameters are designed to extract the mapping relationship between the temperatures of the three-phase windings under different operating conditions; then, the random forest (Random Forest, RF) model is trained to use other monitoring quantities to predict the three-phase winding temperature; in the online anomaly detection stage, if the temperature monitoring data of a phase winding is missing, the extracted three-phase winding temperature mapping relationship is used to fill the missing temperature value; further, the predicted temperature of the previous moment model is used as the input of the RF model at the next moment to obtain the three-phase winding temperature prediction value; finally, the degree of deviation between the actual monitoring value of the three-phase winding temperature and the temperature value predicted by the RF model is calculated to give the anomaly detection result.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A dry-type transformer anomaly detection method based on operating condition identification is proposed. The method can be divided into two stages: offline model training and online anomaly detection.

[0008] (I) Model offline training mainly includes the following steps:

[0009] Step 1) Dry-type transformer historical data collection and preprocessing, the process is as follows:

[0010] Obtain various original historical monitoring data of dry-type transformers through the dry-type transformer monitoring platform, and perform mean filling on the missing values ​​of the original historical monitoring data to obtain the historical monitoring data set Φ of dry-type transformers under normal operating conditions;

[0011] Step 2) Identification and classification of different operating conditions of dry-type transformers. The process is as follows:

[0012] 2.1) Active power and ambient temperature in the historical monitoring data set Φ are selected as features for working condition classification. The active power and ambient temperature after maximum and minimum normalization preprocessing are clustered by the k-means clustering algorithm, in which the Soft-Dynamic Time Warping (Soft-DTW) distance is used to measure the similarity between different curves. The initial number of clusters K is set min , the maximum number of clusters K max , initial γ min value, maximum γmax After each clustering is completed by the k-means algorithm, the Davidson-Bouldin Index (DBI) of this clustering is calculated;

[0013] 2.2) Select the number of clusters and γ value corresponding to the minimum DBI value as the optimal clustering parameter, and use the active power P of the dry-type transformer and the ambient temperature T under the optimal clustering parameter as the optimal clustering parameter. e The clustering results are uniformly coded and used as operating condition labels to divide the historical monitoring data set;

[0014] Step 3) extract the mapping relationship of the temperature between the three-phase windings of the dry-type transformer under each operating condition. The process is as follows:

[0015] 3.1) According to the fluctuation of winding temperature under each operating condition, the number of clusters M and initialization weight w of the SOM clustering algorithm are calculated. j , and set other initialization parameters of SOM algorithm η0, δ0, t max , perform SOM clustering on the three-phase winding temperature data of the dry-type transformer under each operating condition, and obtain M cluster centers under each operating condition;

[0016] Step 4) Establish a prediction model for the three-phase winding temperature of the dry-type transformer under each operating condition

[0017] 4.1) The three-phase current I under each working condition in the historical monitoring data set a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e , and active power P are constructed separately, and the constructed features include root mean square, mean, peak value, and variance. Then, a sliding window processing is performed on the time series to extract the correlation information of each feature time series;

[0018] 4.2) The historical monitoring data after feature construction is selected as the input variable, and the three-phase winding temperature at the corresponding time point is used as the prediction variable. RF is used to establish a prediction model for other monitoring variables and three-phase winding temperature under each working condition.

[0019] 4.3) Feature screening is performed based on the importance of each input feature after random forest training. Some less important features are removed and retrained. Mean absolute error (MAE) and mean square error (MSE) are used as evaluation indicators to test the accuracy of random forest predictions. The maximum absolute error between the predicted value and the actual value under each working condition is used as the upper threshold Ψ for anomaly detection.

[0020] (II) Online anomaly detection mainly includes the following steps:

[0021] Step 5) Calculate the degree of deviation between the three-phase temperature and the predicted value during the online detection process and give the abnormality detection score. The process is as follows:

[0022] 5.1) Calculate the Soft-DTW distance between the active power and ambient temperature in the dataset to be tested and the cluster centers of active power and ambient temperature under several different working conditions. The working condition with the closest distance is the working condition category of the dataset to be tested.

[0023] 5.2) Determine whether there are missing values ​​in the actual operating three-phase temperature in the data set to be tested. If so, use the existing winding temperature to fill the missing value through the mapping relationship established in the above step 3).

[0024] 5.3) The three-phase current I in the data set to be detected a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e The active power P is constructed with the same features as the training set and input into the RF model to obtain the predicted value of the three-phase winding temperature at each monitoring time point. The three-phase temperature output by the model at each moment is used as the input value of the model at the next moment.

[0025] 5.4) Calculate the absolute error between the predicted three-phase winding temperature and the actual three-phase temperature, and determine whether the absolute error exceeds the upper limit threshold Ψ for abnormal detection. If so, it is determined that the three-phase winding temperature of the dry-type transformer in the test data set is abnormal.

[0026] Furthermore, in step 2), the active power P and the ambient temperature T e The formula for unified coding of clustering results is:

[0027] C lable [m]=2*P lable [m]+Te lable [m] (1)

[0028] Among them, P lable [m] and Te lable [m] is the active power P and ambient temperature T on the mth day e Clustering result label, C lable [m] is the operating condition label obtained on the mth day.

[0029] Furthermore, in step 3), the number of clusters M and the initialization weights of the SOM clustering algorithm are The calculation method is:

[0030] 3.1.1) Initialize the number of clusters M; calculate the difference between the maximum and minimum values ​​of the three-phase winding temperature under each operating condition and round it up as the number of clusters in the SOM cluster under that condition, that is, the number of neurons in the output layer:

[0031]

[0032]

[0033] Among them, t aN , t bN , t cN Represents the three-phase temperature values ​​of a, b, and c at the Nth point under this working condition, is the rounding function, max(T a ,T b ,T c )、min(T a ,T b ,T c ) represents the maximum and minimum values ​​of the three-phase temperature under this working condition;

[0034] 3.1.2) Initialize M cluster center neurons and initialize weights w j :

[0035] w j =(t aj ,t bj ,t cj ) (4)

[0036]

[0037] Among them, max(T x )、min(T x ) represents the maximum and minimum values ​​of the x-phase temperature under this working condition, j=1,2,...M, x=a,b,c.

[0038] Furthermore, in step 5.2), the method for filling in the missing winding temperature data is:

[0039] 5.2.1) If the α phase winding temperature data is missing, the missing temperature value is Tα missing , the other two phases are β phase and λ phase, and the temperature value is Tβ existing , Tλ existing , calculate Tβ existing and Tλ existing The Euclidean distance to the corresponding phase temperature in all BMUs under this condition is selected as the Euclidean distance d fill The minimum BMU α phase temperature is used to fill the missing value:

[0040]

[0041] Among them, Tβ i ,Tλ i is the temperature value of the corresponding phase in the BMU under this working condition.

[0042] The advantages of the present invention are:

[0043] 1) A method for dividing the historical data of dry-type transformers into data sets based on operating conditions using their active power and ambient temperature was proposed. A three-phase winding temperature prediction model was established under each operating condition, which improved the applicability and accuracy of temperature prediction.

[0044] 2) The SOM clustering model under each operating condition was optimized, and the mapping relationship between the temperatures of the three-phase windings under each operating condition was more effectively extracted, so that the online detection stage can complete the anomaly detection task when the temperature data of a certain phase is missing.

[0045] 3) By using the model output temperature at the previous moment as the input at the next moment when predicting the dry-type transformer winding temperature, the output error caused by the abnormal actual monitoring winding temperature is reduced, and the anomaly detection performance of the model is improved.

[0046] 4) This method effectively completes the anomaly detection task in the absence of abnormal training data by judging the residual threshold. The theoretical framework of the detection method is simple, easy to implement and has good generalization ability.

[0047] 5) This method does not require prior knowledge of the dry-type transformer design parameters and can perform anomaly detection when the design parameters are completely unknown, which facilitates the promotion of this method in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of a dry-type transformer anomaly detection method based on working condition identification;

[0049] Figure 2 This is a diagram of some monitoring data under normal operating conditions of dry-type transformers;

[0050] Figure 3A diagram showing the eight operating conditions of historical monitoring data for dry-type transformers;

[0051] Figure 4 The residual diagram of the predicted and actual values ​​of the dry-type transformer winding temperature;

[0052] Figure 5 This is the absolute error diagram of the predicted value and actual value of the dry-type transformer winding temperature at different abnormal levels;

[0053] Figure 6 This is the result diagram of using BMU to fill in the missing temperature of the B-phase winding;

[0054] Figure 7 is the absolute error of different abnormal temperature rises after using BMU to fill the B phase temperature data. DETAILED DESCRIPTION

[0055] In order to make the technical solutions and design ideas of the present invention clearer, a detailed description is given below with reference to the accompanying drawings.

[0056] Reference Figure 1-Figure 7 A method for detecting anomalies of a dry-type transformer based on operating condition identification is provided, the method comprising the following steps:

[0057] Step 1) Collection and preprocessing of historical monitoring data of dry-type transformers. The process is as follows:

[0058] Obtain various original historical monitoring data of dry-type transformers through the dry-type transformer monitoring platform. The monitoring data types include three-phase current I a ,I b ,I c , three-phase winding temperature T a ,T b ,T c , active power P, ambient temperature T e 、Ambient humidity H e , the sampling period is 5min, and the sampling time length is one year;

[0059] The actual collected data may be partially or completely missing due to communication errors, gateway disconnection, equipment power outages, etc., so it is necessary to fill in the missing values ​​of the original historical monitoring data. Considering that the operating status of dry-type transformers has obvious periodicity and continuity, data preprocessing is performed here on a daily basis. If there are fewer than 5 missing points in a day, the missing values ​​are filled with the average of the monitoring data for that day. In other cases, the data of that day is removed from the historical monitoring data set to obtain a historical monitoring data set for dry-type transformers under normal operating conditions;

[0060] Step 2) Identification and classification of different operating conditions of dry-type transformers. The process is as follows:

[0061] 2.1) Select active power P and ambient temperature T from the historical monitoring data set e As characteristic parameters for working condition classification, they are normalized to their maximum and minimum values. The k-means clustering algorithm is used to cluster the curves, and the Soft-DTW distance is used to measure the similarity between different curves. The algorithm's clustering process for active power and ambient temperature is similar, and the following uses the clustering process for active power as an example.

[0062] The active power data to be clustered are as follows:

[0063]

[0064] Among them, P m It represents the active power curve of the dry-type transformer on the mth day, p mn represents the active power value at the nth moment on the mth day. The clustering process is as follows:

[0065] 2.1.1) Set the initial number of clusters K min =2, maximum number of clusters K max =6, γ initial value γ min =0.1, γ maximum value γ max =0.9, the maximum number of iterations is 1000;

[0066] 2.1.2) From P1...P m In [1], K active power curves are randomly selected as the centers of clusters, where K is the number of clusters currently set;

[0067] 2.1.3) Calculate the Soft-DTW distance between each time series vector and each cluster center point. The Soft-DTW algorithm distance calculation method is defined as:

[0068]

[0069]

[0070] Among them, R(p in ,p jn ) represents P i 、P j The two active power curves start from the first point and end at P i The nth point, P j The cumulative distance of the best path of the nth point; dist(p in ,p jn ) represents P i The nth point, P j The Euclidean distance between the nth points, the hyperparameter γ can control the smoothness of the result matrix, a i minγ {a1,...,a n}, i=(1,2,...,n);

[0071] 2.1.4) Update the cluster center. For each cluster, calculate the average vector of all time series vectors in the cluster as the new cluster center.

[0072] 2.1.5) Repeat steps 2.1.3)-2.1.4) until the cluster center no longer changes or the maximum number of iterations is reached;

[0073] The algorithm is based on K min The loop algorithm starts with the initial number of clusters, and the number of clusters increases by one after each loop until the number of clusters reaches K max After each clustering is completed by the k-means algorithm, the DBI value of this clustering is calculated; the DBI calculation formula is as follows:

[0074]

[0075] Among them, ω α and ω β Represent the cluster centers of the αth and βth class samples respectively, ||ω α -ω β ||2 represents the distance between two cluster centers, which can measure the degree of dispersion between cluster categories; and Respectively represent the average distances of the αth and βth class samples to their cluster centers, which are used to measure the closeness of samples in each class;

[0076] R α,β Indicates the similarity between the αth class and the βth class:

[0077]

[0078] Then we get the maximum similarity R between the αth class and other classes. α :

[0079]

[0080] The smaller the DBI value, the closer the samples within each class are, the more dispersed the samples between classes are, and the better the clustering effect is;

[0081] 2.2) The number of clusters K and γ corresponding to the minimum DBI value are selected as the optimal clustering parameters to cluster the active power and ambient temperature. The clustering results of daily active power and ambient temperature data are obtained, and the clustering results are uniformly coded as the operating condition label of the day:

[0082] C lable [m]=2*Plable [m]+Te lable [m] (7)

[0083] Among them, P lable [m] and Te lable [m] is the clustering result label of active power and ambient temperature on the mth day, C lable [m] is the operating condition label of the mth day;

[0084] Dividing the historical operation data set into operation data sets under different operation conditions according to the operation condition label;

[0085] Step 3) extract the mapping relationship of the three-phase winding temperature of the dry-type transformer under each operating condition. The process is as follows:

[0086] 3.1) Set the initialization parameters of the SOM clustering model

[0087] Assume that the input layer of SOM is a 3D vector X = (x1, x2, x3), and the output layer is a 2D network with M nodes;

[0088] 3.1.1) Initialize the number of clusters M; calculate the difference between the maximum and minimum values ​​of the three-phase winding temperature under each operating condition and round it up as the number of clusters in the SOM cluster under that condition, that is, the number of neurons in the output layer:

[0089]

[0090]

[0091] Among them, t aN , t bN , t cN Represents the three-phase temperature values ​​of a, b, and c at the Nth point under this working condition, is the rounding function, max(T a ,T b ,T c )、min(T a ,T b ,T c ) represents the maximum and minimum values ​​of the three-phase temperature under this working condition;

[0092] 3.1.2) Initialize M cluster center neurons and initialize weights w j :

[0093] w j =(t aj ,t bj ,t cj ) (10)

[0094]

[0095] Among them, max(T x )、min(T x ) represents the maximum and minimum values ​​of the x-phase temperature under this condition, j = 1, 2, ... M, x = a, b, c;

[0096] Set the initial learning rate of the SOM clustering algorithm to η0 = 0.5, δ0 = 1, t max =10000;

[0097] 3.2) Clustering the three-phase winding temperature using the SOM model

[0098] 3.2.1) Initialize the SOM according to the initialization parameters in step 3.1);

[0099] 3.2.2) Randomly take an input from the training set and normalize it to get

[0100] 3.2.3) Use the following formula to calculate the Euclidean distance d between the input vector and the output node j , find the winning node with the smallest distance:

[0101]

[0102] 3.2.4) Adjust the weights of all nodes in the neighborhood of the winning node. This is shown in the following formula:

[0103] w ij (t+1)=w ij (t)+η(t)g ij [x i -w ij (t)] (13)

[0104]

[0105]

[0106] Among them, t is the current iteration number, t max Represents the maximum number of iterations, x i represent The i-th value of w ij (t) represents the weight of the ith input neuron and the jth winning neuron in the tth iteration, η(t) is the learning rate function, (c x ,c y ) is the coordinate of the winning node in the tth iteration, g ij is the weight adjustment function between the i-th neuron and the winning neuron j in the neighborhood, and δ0 is the update range coefficient;

[0107] 3.2.5) When η(t)≤η min When , the training ends; if the end condition is not met, repeat steps 3.2.2)-3.2.4);

[0108] By performing SOM clustering on the three-phase winding temperature data of the dry-type transformer under each operating condition, M cluster centers under each operating condition are obtained, namely M best matching units (BMUs), which represent the mapping relationship between the three-phase temperatures at different temperature levels under this operating condition.

[0109]

[0110] Among them, BMU M is the Mth cluster center output by the model, t aM ,t bM ,t cM is the three-phase winding temperature value of the Mth cluster center;

[0111] Step 4) Establish a prediction model for the three-phase winding temperature of the dry-type transformer under each operating condition

[0112] 4.1) The three-phase current I under each working condition in the historical monitoring data set a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e , active power P, etc., and construct features including root mean square, mean, peak value, and variance. A sliding window process is performed on the time series with a sliding window size of 6.

[0113] 4.2) Select the three-phase current I after feature construction a ,I b ,I c 、Ambient temperature T e 、Ambient humidity H e , active power P as input variable, the three-phase winding temperature T at the corresponding time point a ,T b ,T c As output variables, RF is used to establish prediction models for other monitoring quantities and three-phase winding temperature under each working condition;

[0114] 4.2.1) The random forest prediction model is:

[0115] {h(L,Θ d ),d=1,2,...,D} (17)

[0116] Among them, h(L,Θ d ) represents the d-th decision tree Θ that constitutes the random forest, L is the input vector of the decision tree, each Θ is independently distributed, representing the random process of extracting the sample data of the d-th tree in the random forest and growing the decision tree;

[0117] When making predictions, the final prediction result y can be obtained based on the output of all decision trees in the model. p :

[0118]

[0119] Among them, D represents the number of trees in RF, y pd is the prediction result of the d-th tree;

[0120] 4.3) Feature screening is performed based on the importance of each input parameter after random forest training. Some less important features are removed and retrained. MAE and MSE are used as evaluation indicators to test the accuracy of random forest predictions. The maximum absolute error between the predicted value and the actual value under each working condition is used as the upper threshold Ψ for anomaly detection:

[0121]

[0122]

[0123]

[0124] Among them, n t is the number of predicted values ​​(n t =1,2,...,N t ), For nth t The predicted temperature value for each point, For nth t The actual temperature value of each point, Ψ is the upper limit of the absolute error of anomaly detection;

[0125] Step 5) Calculate the degree of deviation between the three-phase temperature and the predicted value during the online detection process and give the abnormal detection result. The process is as follows

[0126] 5.1) Input the data set to be tested and calculate the ambient temperature T in the data set to be tested e , active power P and the Soft-DTW distance between the active power and ambient temperature cluster centers under several different working conditions, and select the distance from the ambient temperature T on the day e The cluster center label closest to the active power P is used as the label of the day, and the final working condition label is calculated;

[0127] 5.2) Determine whether there are missing values ​​in the actual operating three-phase temperature in the test data set. If so, fill in the missing values ​​according to the mapping relationship established in step 3);

[0128] 5.2.1) If the α phase winding temperature data is missing, the missing temperature value is Tα missing , the other two phases are β phase and λ phase, and the temperature value is Tβ existing , Tλ existing , calculate Tβ existing and Tλ existing The Euclidean distance to the corresponding phase temperature in all BMUs under this condition is selected as the Euclidean distance d fill The missing value is filled by the minimum α-phase temperature in the BMU.

[0129]

[0130] Among them, Tβ i ,Tλ i is the temperature value of the corresponding phase in the BMU under this working condition;

[0131] 5.3) The three-phase current I in the data set to be detected a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e The active power P is constructed with the same features as the training set and input into the RF model to obtain the three-phase temperature prediction value at each monitoring time point. For each time point, the three-phase winding temperature prediction value output by the model will be used as the input value of the model at the next moment;

[0132] 5.4) Calculate the absolute error between the predicted three-phase winding temperature and the actual three-phase temperature, and determine whether the absolute error exceeds the upper threshold value Ψ for abnormal detection. If so, determine that the three-phase winding temperature of the dry-type transformer in the test data set is abnormal;

[0133] The dry-type transformer monitoring data collected from the real power management system was selected for the experiment. Some of the original data are as follows: Figure 2 The method is used to divide the historical monitoring data of dry-type transformers into 8 data sets under different operating conditions. Figure 3 As shown, a temperature prediction model for dry-type transformers under corresponding operating conditions is established. Figure 4 The residuals of the three-phase temperature prediction under normal operating conditions are shown. Based on this result, the absolute error threshold for abnormal temperature rise of the dry-type transformer winding can be set to 2.5. Temperature abnormality is considered when the threshold is exceeded. Figure 5Taking the B-phase temperature as an example, the absolute error between the predicted and actual winding temperature values ​​under normal operation and abnormality levels of 5% to 15% is shown. It can be seen that when the abnormality level of the winding temperature gradually increases, the absolute error has a more obvious abnormal increasing trend. Figure 6 To simulate the result of using BMU to fill in the B phase temperature when the B phase temperature is missing during the online detection stage, Figure 7 is the absolute error of different abnormal temperature rises after using BMU to fill the B phase temperature data.

[0134] The experimental results show that the proposed method can accurately detect abnormal temperature rises in dry-type transformer windings. Furthermore, as the degree of winding temperature anomaly increases, the absolute error between the predicted and actual winding temperature values ​​increases. This indicates that the predicted temperature values ​​are not affected by the abnormal increase in actual temperature, thus verifying the effectiveness of the proposed method for anomaly detection.

Claims

1. A dry-type transformer abnormality detection method based on working condition identification, characterized in that: The method may include the following steps: 1) Dry-type transformer historical data collection and preprocessing: Get the three-phase current I of the dry-type transformer a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e , active power P historical data, for the missing values ​​in the historical data, the mean value of the data of that day is used to fill in the missing values, and the historical monitoring data set Φ under the normal operating state of the dry-type transformer is obtained; 2) Identification and classification of different operating conditions of dry-type transformers: Select the active power P and ambient temperature T in the historical monitoring data set Φ e As the characteristics of working condition classification; for active power P and ambient temperature T e Curve clustering is performed using the k-means clustering algorithm based on Soft-Dynamic Time Warping Distance, and the initial number of clusters K is set. min , the maximum number of clusters K max , initial smoothness γ min , maximum smoothness γ max After each clustering is completed by the k-means algorithm, the Davidson-Boulding index DBI of this cluster is calculated, and the active power P and ambient temperature T with the smallest DBI are selected. e The clustering results are uniformly coded, and the historical monitoring data set of dry-type transformers is divided based on working conditions; 3) Extract the mapping relationship between the three-phase winding temperatures of the dry-type transformer under each operating condition: According to the fluctuation of winding temperature under each operating condition, the number of clusters M and initialization weight of the self-organizing map network SOM clustering algorithm are calculated. And set other initialization parameters of SOM algorithm η0, δ0, t max Cluster the three-phase winding temperature data of the dry-type transformer under each operating condition to obtain M cluster centers under each operating condition, that is, M best matching units (BMUs), which represent the mapping relationship between the three-phase temperatures at different temperature levels under the operating condition; 4) Establish a prediction model for the three-phase winding temperature of the dry-type transformer under each operating condition: 4.1) The three-phase current I under each working condition in the historical monitoring data set a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e , and active power P are constructed separately. The features constructed for each monitoring quantity include root mean square, mean, peak value, and variance, and the sliding window processing of the time series is performed to extract the correlation information on each feature time series; 4.2) Select historical monitoring data after feature construction as input variables and the three-phase winding temperature at the corresponding time point as the prediction variable. Build a random forest (RF) prediction model for other monitoring variables and the three-phase winding temperature under each operating condition. Perform feature screening based on the importance of each input feature to the results after RF model training. Remove features whose importance does not reach the preset threshold and then retrain. Use mean absolute error (MAE) and mean square error (MSE) as evaluation metrics to test the accuracy of RF model predictions. The maximum absolute error between the predicted value and the actual value under each operating condition is used as the upper threshold Ψ for anomaly detection. 5) Calculate the degree of deviation between the three-phase temperature and the predicted value during the online detection process and give the abnormality detection score: 5.1) Calculate the active power P and ambient temperature T in the data set to be tested e The active power P and ambient temperature T under each working condition obtained in step 2) e The Soft-DTW distance of the cluster center, the closest working condition is the working condition category of the dataset to be tested; 5.2) Determine whether there are missing values ​​in the actual operating three-phase temperature in the test data set. If so, use the existing winding temperature to fill the missing value through the mapping relationship established in step 3) above; 5.3) The three-phase current I in the data set to be detected a ,I b ,I c , three-phase winding temperature T a ,T b ,T c 、Ambient temperature T e 、Ambient humidity H e , the active power P is constructed with the same features as the training set, and the RF model trained in step 4) is input to obtain the predicted value of the three-phase winding temperature at each monitoring time point. The three-phase temperature output by the model at the current moment is used as the input value of the model at the next moment; 5.4) Calculate the absolute error between the predicted three-phase winding temperature and the actual three-phase temperature, and determine whether the absolute error exceeds the upper limit threshold Ψ for abnormal detection. If so, it is determined that the three-phase winding temperature of the dry-type transformer in the test data set is abnormal.

2. The dry-type transformer abnormality detection method based on working condition identification according to claim 1 is characterized in that: In step 2), the active power P and the ambient temperature T e The formula for unified coding of clustering results is: C lable [m]=2*P lable [m]+Te lable [m] (1) Among them, P lable [m] and Te lable [m] is the active power P and ambient temperature T on the mth day e Clustering result label, C lable [m] is the operating condition label obtained on the mth day.

3. The dry-type transformer abnormality detection method based on working condition identification according to claim 1 is characterized in that: In step 3), the number of clusters M and the initialization weight of the SOM clustering algorithm are The calculation method is: 3.1.1) Initialize the number of clusters M; calculate the difference between the maximum and minimum values ​​of the three-phase winding temperature under each operating condition and round it up as the number of clusters in the SOM cluster under that condition, that is, the number of neurons in the output layer: Among them, t aN , t bN , t cN Represents the three-phase temperature values ​​of a, b, and c at the Nth point under this working condition, is the rounding function, max(T a ,T b ,T c )、min(T a ,T b ,T c ) represents the maximum and minimum values ​​of the three-phase temperature under this working condition; 3.1.2) Initialize M cluster center neurons and initialize weights w j : w j =(t aj ,t bj ,t cj ) (4) Among them, max(T x )、min(T x ) represents the maximum and minimum values ​​of the x-phase temperature under this working condition, j=1,2,...M, x=a,b,c.

4. The dry-type transformer abnormality detection method based on working condition identification according to claim 1 is characterized in that: In step 5.2), the method for filling in the missing winding temperature data is: 5.2.1) If the α phase winding temperature data is missing, the missing temperature value is Tα missing , the other two phases are β phase and λ phase, and the temperature value is Tβ existing , Tλ existing , calculate Tβ existing and Tλ existing The Euclidean distance to the corresponding phase temperature in all BMUs under this condition is selected as the Euclidean distance d fill The minimum BMU α phase temperature is used to fill the missing value: in, Tβ i ,Tλ i is the temperature value of the corresponding phase in the BMU under this working condition.

Images