Feature Engineering and Classification Models for Partial Discharge in Power Transformers

2025-05-06 0 0 1.36MB 7 页 10玖币

侵权投诉

Feature Engineering and Classification Models for

Partial Discharge in Power Transformers

Jonathan Wang

Rice University

Houston, Texas

jw96@rice.edu

Kesheng Wu, Alex Sim

Lawrence Berkeley National

Laboratory

Berkeley, California

kwu@lbl.gov,asim@lbl.gov

Seongwook Hwangbo

Hyundai Electric & Energy

Systems Co., Ltd.

Yongin, Korea

smhwangbo@hyundai-electric.

com

ABSTRACT

To ensure reliability, power transformers are monitored for

partial discharge (PD) events, which are symptoms of trans-

former failure. Since failures can have catastrophic cascading

consequences, it is critical to preempt them as early as possi-

ble. Our goal is to classify PDs as corona,oating,particle, or

void, to gain an understanding of the failure location.

Using phase resolved PD signal data, we create a small set

of features, which can be used to classify PDs with high accu-

racy. This set of features consists of the total magnitude, the

maximum magnitude, and the length of the longest empty

band. These features represent the entire signal and not just

a single phase, so the feature set has a xed size and is easily

comprehensible. With both Random Forest and SVM clas-

sication methods, we attain a 99% classication accuracy,

which is signicantly higher than classication using phase

based feature sets such as phase magnitude. Furthermore, we

develop a stacking ensemble to combine several classica-

tion models, resulting in a superior model that outperforms

existing methods in both accuracy and variance.

ACM Reference Format:

Jonathan Wang, Kesheng Wu, Alex Sim, and Seongwook Hwangbo.

2022. Feature Engineering and Classication Models for Partial

Discharge in Power Transformers. In Proceedings of Published at

ICML Workshop for Deep Learning for Safety-Critical in Engineering

Systems (DISE@ICML). ACM, New York, NY, USA, 7 pages. https:

//doi.org/10.1145/nnnnnnn.nnnnnnn

Permission to make digital or hard copies of all or part of this work for

personal or classroom use is granted without fee provided that copies are not

made or distributed for prot or commercial advantage and that copies bear

this notice and the full citation on the rst page. Copyrights for components

of this work owned by others than ACM must be honored. Abstracting with

credit is permitted. To copy otherwise, or republish, to post on servers or to

redistribute to lists, requires prior specic permission and/or a fee. Request

permissions from permissions@acm.org.

DISE@ICML, July 2018, Stockholm, Sweden

ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00

https://doi.org/10.1145/nnnnnnn.nnnnnnn

1 INTRODUCTION

Power transformers are a key element of electric power in-

frastructures. While they have become more reliable, trans-

formers are still susceptible to failure, which has severe

consequences for both operators and users. To detect and

prevent such failures within the large and complex power

transformers, extensive online diagnostic systems have been

developed [

]. This work focuses on analyzing data produced

from one such systems to gain information about the failure.

Insulation failure is the most frequent cause of transformer

failure [

]. Weakness in the insulation system makes trans-

formers susceptible to external events such as lightning

strikes, switching transients, and short-circuits. If the trans-

former insulation degrades to the point that it cannot with-

stand system events such as short-circuit faults or transient

over-voltages [

], an internal arcing event known as partial

discharge (PD) can occur. As such, detecting PD events would

alert transformer operators of imminent transformer failure.

Such measures could also protect other equipment connected

to the transformers, such as Gas Insulation Switchgear (GIS)

and switchboards in the substation, which are also expensive

components of the electric power grid.

Certain types of PD are correlated with dierent parts of

the transformer. For example, in some transformers, corona

PDs are located in the transformer bushing or insulation

material. Therefore, determining the type of PD provides

a rough location for the PD source. Machine classication

is a very useful way to resolve transformer problems early

on, and can be used by transformer operators to determine

whether to halt transformer operation, all without the need

for careful examination by engineers. By classifying the PD,

we narrow down the location of the PD. We can then install

UHF sensors around the rough position to collect PD signals

to identify the precise position of the PD for repair.

In this paper, we consider four types of PDs - corona,oat-

ing,particle, and void, with the goal of analyzing PD signal

data to identify what type of PD is present. The data we

examine is phase resolved, meaning it is divided into cy-

cles of a number of phases. Our data samples consist of 3840

arXiv:2210.12216v1 [cs.LG] 21 Oct 2022

DISE@ICML, July 2018, Stockholm, Sweden Jonathan Wang, Kesheng Wu, Alex Sim, and Seongwook Hwangbo

points divided into 60 cycles of 64 phases. We examine actual

transformer data and test several feature sets and classica-

tion methods to classify PD events with high accuracy. The

contributions of our work are as follows:

•

Develop a set of

meta-features

(total magnitude, max-

imum magnitude, and the length of the longest empty

band) which are more comprehensible than standard

features.

•

Test models such as Logistic Regression and Random

Forest for PD classication, and combine them to pro-

duce an

ensemble model

that performs better than

any single classication method.

2 RELATED WORKS

This work focuses on classifying PDs based on voltage data.

These four PD types occur in power equipment such as trans-

formers and GIS (Gas Insulation Switchgear). In order to clas-

sify and analyze the characteristics of the PD types, many

experiments have been done with GIS, which has a simpler

structure than transformers.

In [

], acoustic methods are proposed to detect corona PDs

in live parts of GIS. By analyzing the eect on the particle

motion and discharge characteristics in the GIS, the discharge

characteristic spectrum of linear particles with tip corona was

obtained, and used widely for particle pattern recognition

[

]. In [

], ve dierent types of defects were implemented

articially in transformer models to investigate the resulting

PD signal characteristics. However, there was a limitation as

dierent transformers can have dierent failures depending

on the transformer structure.

Unlike the above experiments, which involve manually

examining signal characteristics, machine learning based PD

classication methods consist of extracting features from

the data and training models on those features. There are

several existing feature sets such as statistical parameters or

fractal features [

] or partial power [

]. Some of the more

common feature sets are phase magnitude, which consists of

information regarding the magnitude of each phase of the

data [3, 10] and discrete wavelet transform [3, 4, 10].

Since we are working with actual transformer data as

opposed to simulated data as in the above works, we im-

plement methods to reduce data noise. We also present a

smaller feature set of xed size to represent the PD signal

data, signicantly simplifying the model and making the

model more comprehensible.

For the classication model, the primary methods that

have been tested are SVM [

] and Neural Networks

[

]. A variation of SVM, Fuzzy SVM (FSVM), was

also explored in [

]. FSVM allows for fuzzy membership in

classes to resolve unclassiable regions [

] by weighting

the samples based on distance from the class center [

]. In

(a) Corona PD (b) Floating PD

Figure 1: Heatmaps of Sample PDs

addition to these methods, we also experiment with Random

Forest, Logistic Regression, and Gradient Boosting. In addi-

tion, we combine these methods using stacking [

] to create

an ensemble that utilizes the strengths of each model.

3 METHODS

The goal of our work is to classify four types of PD - corona,

oating,particle, and void. We accomplish this in two steps:

•Extract features to represent signal data

•Train machine learning model on features

Our data is 328 PD signals gathered by the transformer

sensors labelled as 85 corona, 99 oating, 80 particle, and

64 void. Each data sample contains 3840 magnitude points

over one second. These points are broken up into 60 cycles

of 64 phases. Figure 1 shows heatmaps of each type of PD.

The x and y axes indicate the cycle and phase and the color

indicates the magnitude at that time.

Feature Engineering

Phase Magnitude. We notice that there are clear patterns

along the phases of each data sample. For instance the corona

PD has a single thick band while the particle PD is scattered

lightly across most of the phases. Thus, we compute the total

phase magnitude for each type of PD. The phase magnitude

is the sum of the magnitudes of each phase, given by

𝑚𝑖=



𝑗=1

𝑀𝑖 𝑗

𝑚𝑝={𝑚1, ..., 𝑚64 }

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

FeatureEngineeringandClassificationModelsforPartialDischargeinPowerTransformersJonathanWangRiceUniversityHouston,Texasjw96@rice.eduKeshengWu,AlexSimLawrenceBerkeleyNationalLaboratoryBerkeley,Californiakwu@lbl.gov,asim@lbl.govSeongwookHwangboHyundaiElectric&EnergySystemsCo.,Ltd.Yongin,Koreasmhwangbo@...

收起<<

Feature Engineering and Classification Models for Partial Discharge in Power Transformers.pdf

共7页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Feature Engineering and Classification Models for Partial Discharge in Power Transformers

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: