CN115169704A - Method and device for CVT error state prediction based on incremental ensemble learning model - Google Patents

Abstract

The invention discloses a CVT error state prediction method and device based on an incremental integrated learning model. The method comprises the steps of: dividing the CVT historical data of power failure verification into several data blocks, and generating corresponding data blocks according to the several data blocks. A number of base models, fuse the base model into a reference state prediction model; obtain the first CVT real-time data, and detect whether concept drift occurs between the first CVT real-time data and CVT historical data; When concept drift occurs , obtain the incremental data of the first CVT real-time data relative to the CVT historical data, and generate an incremental base model according to the incremental data; generate an adaptive incremental model according to the incremental base model and the reference state prediction model A quantitative integrated learning model is used to obtain second CVT real-time data, and an error state prediction is performed on the second CVT real-time data according to the adaptive incremental integrated learning model. The invention improves the accuracy of CVT error state prediction.

Description

Method and device for CVT error state prediction based on incremental ensemble learning model

technical field

The invention relates to the technical field of CVT error state prediction, in particular to a CVT error state prediction method and device based on an incremental integrated learning model.

Background technique

The voltage transformer is an important measurement equipment in the power system. Its primary winding is connected to the high-voltage power grid, and the secondary winding is connected to measuring, metering, protection and other devices, and is used to convert the high-voltage signal on the primary side into a small low-voltage signal for secondary equipment. use.

The long-term operation experience shows that due to the increase of the service life of the voltage transformer, there is a certain risk of out-of-tolerance in the voltage transformer after several years of operation.

Continued operation of the out-of-tolerance voltage transformer will bring huge losses to the gateway metering trade settlement of the three parties, and even affect the stable operation of the power system. Therefore, in order to ensure the accuracy of measurement and the safe operation of the power system, it is necessary to evaluate and replace the voltage transformer with abnormal error state in time. There are mature offline evaluation methods for periodic offline evaluation of voltage transformers, but due to the difficulty of non-faulty power outage operations for high-voltage transmission networks, it is difficult for this method to cover all voltage transformers to be verified within a specified period. ; And the environmental electromagnetic field during offline evaluation is different from that during online operation, resulting in a certain deviation between the evaluation results and the actual situation, which further leads to a large number of voltage transformers in operation in the substation that are not inspected over time and have unknown errors.

In order to solve the shortcomings of the periodic off-line evaluation method, the existing technology adopts the on-line evaluation method under the condition of no power outage to realize the real-time on-line monitoring of the error state of the voltage transformer. The existing online evaluation technology is to analyze and process the signals collected by each device in the power system and based on the data-driven principle, so as to evaluate the error state of the voltage transformer. The approximate model relies on a large amount of data and calculations to characterize the real error state of the voltage transformer in real time. However, the existing technology has the following shortcomings: in terms of data, the existing technology does not consider the concept drift in the data set. Concept drift means that the concepts contained in the data set have changed, such as equipment aging, sudden changes in operating conditions, etc. As a result, the concepts contained in the old and new data are no longer consistent. Once the phenomenon of concept drift occurs in the data set, it will affect the accuracy of characterizing the true error state of the voltage transformer based on the data-driven principle.

SUMMARY OF THE INVENTION

The invention provides a CVT error state prediction method and device based on an incremental integrated learning model, which improves the accuracy of CVT error state prediction.

An embodiment of the present invention provides a CVT error state prediction method based on an incremental ensemble learning model, comprising the following steps:

Divide the CVT historical data of power failure verification into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model;

acquiring the first CVT real-time data, and detecting whether concept drift occurs between the first CVT real-time data and the CVT historical data;

When concept drift occurs, acquiring incremental data of the first CVT real-time data relative to the CVT historical data, and generating an incremental base model according to the incremental data;

Generate an adaptive incremental integrated learning model according to the incremental base model and the reference state prediction model, obtain second CVT real-time data, and perform error state prediction on the second CVT real-time data according to the adaptive incremental integrated learning model .

Further, generating a corresponding number of base models according to the several data blocks, and merging the base models into a reference state prediction model, including the following steps:

The k data blocks are correspondingly formed into k first base models, and the cross-validation method is used to update the k first base models into corresponding k second base models; wherein, k is a positive integer greater than 3;

The k second base models are fused into a reference state prediction model.

Further, for any one of the first base models, the first base model is generated according to one data block in the k data blocks, and the remaining k-1 data block pairs except the one data block are used. The first base model is cross-validated to obtain a corresponding second base model.

Further, after the incremental base model is replaced with the base model with the worst classification effect among the k second base models, the incremental base model and the remaining k-1 second base models are fused to obtain an adaptive Incremental ensemble learning models.

Further, after the incremental base model is replaced with the base model with the worst classification effect among the k first base models, the cross-validation method is used to compare the remaining k-1 first base models and the incremental base models. The model is updated to the corresponding k second base models;

The k second base models are fused into a reference state prediction model.

Further, the incremental base model and the reference state prediction model are fused to obtain an adaptive incremental integrated learning model.

Further, calculate KL divergence and drift threshold according to the first CVT real-time data and CVT historical data;

Whether concept drift occurs is determined according to the KL divergence and drift threshold.

Further, mean-shift clustering is performed on the entire CVT data set formed by the CVT historical data and the first CVT real-time data, and the drift threshold is calculated according to the formula R=D-2r, where D is the average value between the cluster centers. distance, r is the average radius of the clusters.

Further, an adversarial neural network algorithm is used to perform oversampling processing on the minority class samples in the incremental data, and the incremental base model is generated according to the incremental data after the oversampling processing.

Another embodiment of the present invention provides a CVT error state prediction device based on an incremental integrated learning model, including a base model generation module, a concept drift detection module, an incremental base model generation module, and an error state prediction module;

The base model generation module is used to equally divide the CVT historical data of the power outage test into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model;

The concept drift detection module is configured to acquire the first CVT real-time data, and detect whether concept drift occurs between the first CVT real-time data and the CVT historical data;

The incremental base model generation module is configured to acquire incremental data of the first CVT real-time data relative to the CVT historical data when concept drift occurs, and to generate an incremental base model according to the incremental data;

The error state prediction module is used to generate an adaptive incremental ensemble learning model according to the incremental base model and the reference state prediction model, obtain the second CVT real-time data, and perform the second CVT real-time data according to the adaptive incremental ensemble learning model. Two CVT real-time data for error state prediction.

The embodiment of the present invention has the following beneficial effects:

The invention provides a CVT error state prediction method and device based on an incremental integrated learning model. The method detects the concept drift phenomenon that may occur in the CVT historical data and the CVT real-time data, and according to the concept drift detection results, adopts adaptive Incremental integrated classification model, in response to dynamic data flow, performs adaptive error state evaluation, improves the accuracy of the error state evaluation model and the adaptability range of the model, and further improves the accuracy of CVT error state evaluation.

Description of drawings

1 is a schematic flowchart of a CVT error state prediction method based on an incremental integrated learning model provided by an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a CVT error state prediction apparatus based on an incremental ensemble learning model provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

As shown in FIG. 1 , a method for predicting a CVT error state based on an incremental ensemble learning model provided by an embodiment of the present invention includes the following steps:

Step S101: Divide the CVT historical data of the power failure verification into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model.

As one of the embodiments, generating a corresponding number of base models according to the several data blocks, and merging the base models into a reference state prediction model, including the following steps:

Step S11: form k first base models corresponding to the k data blocks, and then use the cross-validation method to update the k first base models into corresponding k second base models; wherein, k is a positive value greater than 3. Integer. Specifically, the k first base models are cross-checked for K times to generate k second base models, and for any of the first base models, the first base model is generated according to one data block in the k data blocks. A base model is used, and the remaining k-1 data blocks except the one data block are used to perform cross-validation on the first base model to obtain a corresponding second base model.

Step S12: Integrate the k second base models into a reference state prediction model.

As one of the embodiments, the CVT historical data of power failure verification is divided into k data blocks D={D ₁ , D ₂ , D ₃ , ..., D _k }, where k=1, 2, ..., a; a≥1.

D ^t represents the window data at historical time t, and defines each window data:

Then its underlying data distribution is:

表示特征向量，

表示CVT样本的类标签。本发明选择的CVT的特征向量为组间幅值比差f和组间相位差

CVT样本的类标签为：正常、异常、告警3种状态标签。where n is the number of samples,

represents the feature vector,

Represents the class labels of the CVT samples. The eigenvectors of the CVT selected by the present invention are the amplitude ratio difference f between groups and the phase difference between groups

The class labels of CVT samples are: normal, abnormal, and alarm status labels.

Among them, A, B, C represent the three phases of the CVT, i represents the group, j represents the row of the vector, and V _Aij represents the amplitude of the A-phase CVT in the jth row of the ith group; conceptual drift refers to the change in the underlying data distribution. ,which is:

分布的最优模型M^t，因此，根据公式(6)生成所述第一基模型：At time t, the goal of the ensemble model is to construct a

The optimal model M ^t of the distribution, therefore, the first base model is generated according to formula (6):

The first base model is updated according to the loss function (7):

和真实值

的均方根误差，λ为学习参数(初始值为0)，θ(M_new)为新模型的参数，θ(M_orign)为原始模型的参数。(7); where, χ(·) represents an arbitrary distribution function, E(·) represents the expected value of a random variable, L is the loss function, where k=1, 2, ..., N; MSE represents the predicted value

and true value

The root mean square error of , λ is the learning parameter (the initial value is 0), θ(M_new) is the parameter of the new model, and θ(M_orign) is the parameter of the original model.

Step S102: Acquire first CVT real-time data, and detect whether concept drift occurs between the first CVT real-time data and CVT historical data.

As one of the embodiments, the KL divergence and drift threshold are calculated according to the first CVT real-time data and the CVT historical data; whether concept drift occurs is determined according to the KL divergence and the drift threshold.

Perform mean-shift clustering on the entire CVT data set formed by the CVT historical data and the first CVT real-time data, and calculate the drift threshold according to the formula R=D-2r, where D is the average distance between the cluster centers, and r is the average radius of the clusters.

The KL divergence is calculated according to formula (8):

In the formula, record the CVT historical data as D _old , and its mean vector and covariance matrix are respectively μ _old and η _old ; the first CVT real-time data is D _new , and its mean vector and covariance matrix are respectively μ _new and n _new ; d is the input sample dimension, preferably, d=2; tr represents the trace of the matrix; T represents the transposition of the matrix.

Step S103: When concept drift occurs, acquire incremental data of the first CVT real-time data relative to the CVT historical data, and generate an incremental base model according to the incremental data.

An adversarial neural network algorithm is used to perform oversampling processing on the minority class samples in the incremental data, and the incremental base model is generated according to the incremental data after the oversampling processing.

Step S104: Generate an adaptive incremental integrated learning model according to the incremental base model and the reference state prediction model, obtain second CVT real-time data, and perform the second CVT real-time data according to the adaptive incremental integrated learning model. Error state prediction. The error status includes normal, abnormal and alarm.

As one of the embodiments, after replacing the base model with the worst classification effect among the k second base models by the incremental base model, the incremental base model and the remaining k-1 second base models are replaced by the incremental base model. The fusion results in an adaptive incremental ensemble learning model.

As one of the embodiments, after replacing the incremental base model with the base model with the worst classification effect among the k first base models, the cross-validation method is used to compare the remaining k-1 first base models and all The incremental base models are updated into corresponding k second base models; the k second base models are fused into a reference state prediction model.

As one of the embodiments, the incremental base model and the reference state prediction model are fused to obtain an adaptive incremental integrated learning model.

When the incremental calculation of the incremental base model does not exceed the maximum memory and limited time of the computer, the adaptive incremental integrated learning model is obtained by fusing the incremental base model and the reference state prediction model. When the incremental calculation of the incremental base model exceeds the maximum memory and limited time of the computer, the incremental base model is used to replace the base model, and then the adaptive incremental integrated learning model is obtained by fusion.

The invention detects the concept drift phenomenon that may occur in the CVT historical data and the CVT real-time data, and adopts an adaptive incremental integrated classification model according to the concept drift detection result to deal with the dynamic data flow, and performs adaptive error state evaluation to improve the The error state evaluates the accuracy of the model and the adaptability range of the model, thereby improving the accuracy of the CVT error state evaluation.

On the basis of the above-mentioned embodiments of the invention, the present invention correspondingly provides an embodiment of a device item, as shown in FIG. 2 ;

Another embodiment of the present invention provides a CVT error state prediction device based on an incremental integrated learning model, including a base model generation module 101, a concept drift detection module 102, an incremental base model generation module 103, and an error state prediction module 104;

The base model generation module is used to equally divide the CVT historical data of the power outage test into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model;

The concept drift detection module is configured to acquire the first CVT real-time data, and detect whether concept drift occurs between the first CVT real-time data and the CVT historical data;

The incremental base model generation module is configured to acquire incremental data of the first CVT real-time data relative to the CVT historical data when concept drift occurs, and to generate an incremental base model according to the incremental data;

The error state prediction module is used to generate an adaptive incremental ensemble learning model according to the incremental base model and the reference state prediction model, obtain the second CVT real-time data, and perform the second CVT real-time data according to the adaptive incremental ensemble learning model. Two CVT real-time data for error state prediction.

For the convenience and simplicity of description, the device item embodiments of the present invention include all the implementation manners in the above-mentioned embodiments of the CVT error state prediction method based on the incremental ensemble learning model, which will not be repeated here.

It should be noted that the device embodiments described above are only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical unit, that is, it can be located in one place, or it can be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution in this embodiment. In addition, in the drawings of the apparatus embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative effort.

The above are the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made, and these improvements and modifications may also be regarded as It is the protection scope of the present invention.

Those of ordinary skill in the art can understand that the realization of all or part of the processes in the above embodiments can be accomplished by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. During execution, the processes of the above-mentioned embodiments may be included. The storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM) or the like.

Claims (10)

1. a CVT error state prediction method based on incremental integrated learning model, is characterized in that, comprises the following steps: Divide the CVT historical data of power failure verification into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model; acquiring the first CVT real-time data, and detecting whether concept drift occurs between the first CVT real-time data and the CVT historical data; When concept drift occurs, acquiring incremental data of the first CVT real-time data relative to the CVT historical data, and generating an incremental base model according to the incremental data; Generate an adaptive incremental integrated learning model according to the incremental base model and the reference state prediction model, obtain second CVT real-time data, and perform error state prediction on the second CVT real-time data according to the adaptive incremental integrated learning model . 2. the CVT error state prediction method based on incremental ensemble learning model according to claim 1, is characterized in that, according to described several data blocks, the base model of corresponding quantity is generated, and described base model is fused into a reference state Predictive model, including the following steps: The k data blocks are correspondingly formed into k first base models, and the cross-validation method is used to update the k first base models into corresponding k second base models; wherein, k is a positive integer greater than 3; The k second base models are fused into a reference state prediction model. 3. The CVT error state prediction method based on an incremental integrated learning model according to claim 2, wherein, for any of the first base models, the first base model is generated according to a data block in the k data blocks. A base model is used, and the remaining k-1 data blocks except the one data block are used to perform cross-validation on the first base model to obtain a corresponding second base model. 4. the CVT error state prediction method based on the incremental integrated learning model according to claim 3, is characterized in that, replaces the basic model with the worst classification effect in described k second basic model by described incremental basic model Then, the incremental base model and the remaining k-1 second base models are fused to obtain an adaptive incremental ensemble learning model. 5. the CVT error state prediction method based on the incremental integrated learning model according to claim 3, is characterized in that, replaces the basic model with the worst classification effect in described k first basic model by described incremental basic model Then, the cross-validation method is used to update the remaining k-1 first base models and the incremental base models into corresponding k second base models; The k second base models are fused into a reference state prediction model. 6 . The CVT error state prediction method based on an incremental integrated learning model according to claim 3 , wherein an adaptive incremental integrated learning model is obtained by fusing the incremental base model and the reference state prediction model. 7 . 7. the CVT error state prediction method based on the incremental integrated learning model according to claim 4 or 5 or 6, is characterized in that, calculates KL divergence and drift threshold value according to described first CVT real-time data and CVT historical data; Whether concept drift occurs is determined according to the KL divergence and drift threshold. 8. the CVT error state prediction method based on incremental integrated learning model according to claim 7, is characterized in that, the CVT data set that described CVT historical data and the first CVT real-time data are formed integrally carries out mean shift clustering, The drift threshold is calculated according to the formula R=D-2r, where D is the average distance between cluster centers, and r is the average radius of the clusters. 9. The CVT error state prediction method based on an incremental integrated learning model according to claim 8, wherein an adversarial neural network algorithm is used to perform oversampling processing on the minority samples in the incremental data, and according to the oversampling The processed incremental data generates the incremental base model. 10. A CVT error state prediction device based on an incremental integrated learning model, characterized in that it comprises a base model generation module, a concept drift detection module, an incremental base model generation module and an error state prediction module; The base model generation module is used to equally divide the CVT historical data of the power outage test into several data blocks, generate a corresponding number of base models according to the several data blocks, and fuse the base models into a reference state prediction model; The concept drift detection module is configured to acquire the first CVT real-time data, and detect whether concept drift occurs between the first CVT real-time data and the CVT historical data; The incremental base model generation module is configured to acquire incremental data of the first CVT real-time data relative to the CVT historical data when concept drift occurs, and to generate an incremental base model according to the incremental data; The error state prediction module is used to generate an adaptive incremental ensemble learning model according to the incremental base model and the reference state prediction model, obtain the second CVT real-time data, and perform the second CVT real-time data according to the adaptive incremental ensemble learning model. Two CVT real-time data for error state prediction.

Images