CN113759208A

CN113759208A - Abnormal waveform identification method based on fault indicator

Info

Publication number: CN113759208A
Application number: CN202110616402.4A
Authority: CN
Inventors: 曹乾磊; 王彦萍; 刘军; 李建赛; 狄克松
Original assignee: Qingdao Topscomm Communication Co Ltd
Current assignee: Qingdao Topscomm Communication Co Ltd
Priority date: 2021-06-02
Filing date: 2021-06-02
Publication date: 2021-12-07

Abstract

The invention discloses an abnormal waveform identification method based on a fault indicator, which comprises the following steps: collecting original waveforms uploaded to a main station by a fault indicator and carrying out direct current removal processing on the original waveforms; manually adding classification labels to all collected waveform data; combining the original waveform data and the waveform data after the direct current is removed into new data, dividing the new data and the corresponding labels into a training set and a test set, and feeding the training set and the test set into a TextCNN network model; performing parameter adjustment, training and testing on the network model to obtain an optimal model; and performing direct current removal processing on the waveform to be recognized, and inputting the original data of the waveform to be recognized and the processed data into the optimal model to obtain a recognition result. According to the method, only the original waveform needs to be subjected to direct current removal processing, no further feature extraction is needed, and compared with the traditional CNN model, the TextCNN model can more accurately capture the relation between voltage and current, so that the identification accuracy is improved, and the advantages of simple model and high engineering practicability are achieved on the basis of ensuring the running state of a power grid.

Description

Abnormal waveform identification method based on fault indicator

Technical Field

The invention relates to the technical field of medium-voltage power distribution networks, in particular to an abnormal waveform identification method based on a fault indicator.

Background

The running state of the power distribution network is crucial to the stable running of the power grid, and when a power system fails, the power system needs to be positioned and cleared as soon as possible. The fault indicator (comprising an acquisition unit and a collection unit) is used as a fault indicating and warning device, can automatically record all electric quantities before and after a fault moment and transmit the electric quantities to a main station in a waveform mode, and the power distribution main station analyzes the fault type according to the waveform uploaded by the fault indicator and the change rule of the electric quantities when a power system is grounded or short-circuit fault occurs and judges the fault occurrence position. The fault indicator also has the advantages of no maintenance, low power consumption, charged loading and unloading and the like, and is widely applied to a 10kV power distribution network system based on the advantages. However, due to the complex working environment, problems may occur in the aspects of sampling by the acquisition unit, communication between the acquisition unit and the collection unit, communication between the collection unit and the master station, and the like, and the waveform quality uploaded to the master station by the fault indicator is difficult to guarantee, and may contain many abnormal waveforms, thereby affecting the judgment of the fault position, and causing the occurrence of the conditions of missed judgment and wrong judgment.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides an abnormal waveform identification method based on a fault indicator, which is used for carrying out direct current removal processing on a waveform uploaded to a main station by the fault indicator, combining the waveform with original data and inputting the combined waveform into a trained TextCNN model for calculation so as to realize the identification of the abnormal waveform.

The purpose of the invention can be realized by the following technical scheme:

an abnormal waveform identification method based on a fault indicator comprises the following steps:

step 1: collecting original waveforms uploaded to a main station by a fault indicator and carrying out direct current removal processing on the original waveforms;

step 2: manually adding classification labels to all collected waveform data;

and step 3: combining the original waveform data and the waveform data after the direct current is removed into new data, dividing the new data and the labels corresponding to the new data into a training set and a test set, and feeding the training set and the test set into a TextCNN network model;

and 4, step 4: performing parameter adjustment, training and testing on the network model to obtain an optimal model;

and 5: and performing direct current removal processing on the waveform to be recognized, and inputting the original data of the waveform to be recognized and the processed data into the optimal model to obtain a recognition result.

Further, each group of waveforms of the main station uploaded by the fault indicator in the step 1 includes 8 channel data including three-phase electric field and zero-sequence electric field data and three-phase current and zero-sequence current data, and after performing dc removal processing, 8 channel data are obtained again, and finally 16 channel data are obtained;

further, the dc removal method used in step 1 is a segmented dc removal method, and the calculation formula is as follows:

wherein, data _ new_iFor removing the data of the ith point after DC_iFor the data of the ith point before the direct current is removed, the cyclePoint is the total point number of a cycle;

further, the label classification method in step 2 is as follows: marking the normal waveform as 1, and marking the abnormal waveform as-1;

further, the method used in the step 3 for dividing the training set and the test set is hierarchical sampling;

further, the constraint overfitting method used in the training process in step 4 includes Early Stopping, L1, L2 regularization, and Dropout.

The invention has the beneficial effects that: the original waveform is only required to be subjected to direct current removal processing, waveform characteristics are not required to be further extracted, and compared with the traditional CNN method, the TextCNN method can more accurately capture the relation between voltage and current, the identification accuracy is improved, the method has the advantages of being simple in model and easy to realize on the basis of ensuring the running state of the power grid, and has strong engineering practicability.

Drawings

FIG. 1 is a general flow diagram of the present invention.

FIG. 2 is a diagram illustrating the structure of the TextCNN model according to the present invention.

Fig. 3 is a schematic diagram of dimensional changes of waveform data in a training or testing process according to the present invention.

Fig. 4 is a diagram of a field topology provided by an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the invention.

With reference to fig. 1, a method for identifying an abnormal waveform based on a fault indicator includes the following steps:

step 1: collecting original waveforms uploaded to a main station by a fault indicator and carrying out direct current removal processing on the original waveforms; each group of waveforms of the main station uploaded by the fault indicator comprise 8 channel data including three-phase electric field and zero-sequence electric field data and three-phase current and zero-sequence current data, the 8 channel data are obtained after the direct current removal processing is carried out, and finally 16 channel data are obtained; the used DC removing method is segmented DC removing, and the calculation formula is as follows:

step 2: manually add classification labels to all collected waveform data: marking the normal waveform as 1, and marking the abnormal waveform as-1;

and step 3: combining original waveform data and waveform data subjected to direct current removal into new data, dividing the new data and labels corresponding to the new data into a training set and a test set by a hierarchical sampling method, and feeding the training set and the test set into a TextCNN network model, wherein the structure of the model network is shown in FIG. 2;

and 4, step 4: performing parameter adjustment, training and testing on the network model to obtain an optimal model; the limiting overfitting methods used in the training process include Early Stopping, L1, L2 regularization and Dropout;

and 5: and performing direct current removal processing on the waveform to be recognized, and inputting the original data of the waveform to be recognized and the processed data into the optimal model to obtain a recognition result. The dimension change of the waveform data from input to output is shown in fig. 3, the final output dimension is (number of samples, 2), and the sample identification result is obtained when the probability value in each row is larger.

Fig. 4 shows a field topology diagram, which includes 6 index measurement points, and the classification result obtained by inputting the waveform of each measurement point into the model is as follows:

indicating a measuring point	Model output results	Classification result
			1	(0.95，0.05)	Normal waveform
2	(0.1，0.9)	Abnormal waveform
			3	(0.83，0.17)	Normal waveform
4	(0.76，0.24)	Normal waveform
			5	(0.91，0.09)	Normal waveform
6	(0.22，0.78)	Abnormal waveform

In the embodiment, the identification result of the method is accurate and consistent with the actual result.

The above-mentioned embodiments are illustrative of the specific embodiments of the present invention, and are not restrictive, and those skilled in the relevant art can make various changes and modifications to obtain corresponding equivalent technical solutions without departing from the spirit and scope of the present invention, so that all equivalent technical solutions should be included in the scope of the present invention.

Claims

1. An abnormal waveform identification method based on a fault indicator is characterized by comprising the following steps:

step 2: manually adding classification labels to all collected waveform data;

and step 3: combining the original waveform data and the waveform data after the direct current is removed into new data, dividing the new data and the corresponding labels into a training set and a test set, and feeding the training set and the test set into a TextCNN network model;

2. The method for identifying the abnormal waveform based on the fault indicator according to claim 1, wherein each group of waveforms uploaded to the main station by the fault indicator in the step 1 comprises 8 channel data including three-phase electric field and zero-sequence electric field data and three-phase current and zero-sequence current data, and the 8 channel data are obtained after the dc removal processing is performed, so that 16 channel data are obtained finally;

3. the method for identifying the abnormal waveform based on the fault indicator according to claim 1, wherein the dc removing method used in the step 1 is segmented dc removing, and the calculation formula is as follows:

wherein, data _ new_iFor removing the data of the ith point after DC_iThe total number of points of a cycle of the CyclePoint is the data of the ith point before direct current.

4. The abnormal waveform identification method based on the fault indicator as claimed in claim 1, wherein the label classification method in the step 2 is as follows: marking the normal waveform as 1, and marking the abnormal waveform as-1;

5. the abnormal waveform identification method based on the fault indicator as claimed in claim 1, wherein the method used for dividing the training set and the test set in the step 3 is hierarchical sampling;

6. the fault indicator-based abnormal waveform identification method according to claim 1, wherein the constraint overfitting methods used in the training process in the step 4 comprise Early Stopping, L1, L2 regularization and Dropout.