Sliding-Time-Window_and_Event-Trigger_Based_Data_Collection_Strategy_for_Non-intrusive_Load_Monitoring__GoldenDB.pdf

tinge

6页

0次

2023-10-26

免费下载

Sliding-Time-Window and Event-Trigger Based

Data Collection Strategy for Non-intrusive Load

Monitoring

Jian Yang*, Wenjia Hua, Mengfei Xu

School of Computer

NUPT-JSTZ Carbon Peaking & Carbon Neutrality Application Technology R&D Institute

Nanjing University of Posts and Telecommunications

Nanjing, China

yangj@njupt.edu.cn

Zhiyou Ouyang

Institute of Advanced Technology for Carbon Neutrality

Nanjing University of Posts and Telecommunications

Nanjing, China

netivs@qq.com

Han Zhang

School of Cyber Science and Engineering

Southeast University and ZTE Corporation

Nanjing, China

zhang.han1@zte.com.cn

Abstract—Non-intrusive load monitoring (NILM) technology

can effectively identify the type and operation state of the

electrical equipments in the load, so as to help the better analyze

the power consumption state and power consumption of users,

and achieve better energy saving and emission reduction and

load optimization control. However, the large amount of data

collected by non-invasive load monitoring terminals, especially

high-frequency data, brings challenges to data acquisition, trans-

mission and training of non-intrusive load monitoring models.

Hence, in this paper, a sliding-time-window and event trigger

based non-intrusive load monitoring data acquisition strategy

is proposed, which can select power equipment of state change

events through the event trigger mechanism, and ﬁx the length

of collect data samples and ensure not lose equipment state

change process of context information by using sliding-time-

window. Experiments by using widely used non-intrusive load

monitoring algorithm evaluation toolkit that named NLIMTK,

and widely used public dataset which named REDD, show

that the proposed data collection strategy can not only make

non-intrusive load monitoring data collected by the size of the

largest decreased by 97.07%, and the prediction performance has

improved signiﬁcantly, which fully demonstrate the effectiveness

of the proposed method.

Index Terms—non-intrusive load monitoring, event trigger,

data collection, load control, power system

I. INTRODUCTION

Non-intrusive load monitoring (NILM) is a power mon-

itoring technology, which gathers power parameters from a

power load collection device (usually a smart meter) deployed

This work was supported by the National Natural Science Key Fund

(NO.61533010, No.6183308, No.61933005), Natural Science Fund of Nan-

jing University of Posts and Telecommunications (20190256) and Henan

province science and technology research projects (High and new technology)

(No.182102210106).

* Jian Yang is the corresponding author.

at the entrance of power supply for residents, and identiﬁed

what kinds of appliances users have used and when the user

open or close those appliances uses machine learning and

other methods [1], [2]. Identify the power consumption of

various load devices in the residential user’s room can help the

power company to better understand the consumption habits

of users, so as to provide better electricity experience for

users. This technology is easy to deploy, does not need to

install any equipment in the user’s room, and saves a lot

of manpower and material resources compared with intrusive

load monitoring, so it has been widely concerned [3]. Non-

intrusive load monitoring can identify the user type and oper-

ation of electrical equipment, which help power grid manage-

ment personnel to understand the user’s electricity usage and

power consumption, and then the user type identiﬁcation and

optimize the power equipment control, precision load control,

power system, power distribution network optimal operation

and energy conservation and emissions reduction targets, To

provide customers with higher quality electricity services.

Studies have shown that information on energy consumption

at the appliance level can help reduce energy consumption by

4-15% [4], [5].

Typical application scenario of non-intrusive load monitor-

ing [6] is shown in the following Fig.1, in which, the non-

intrusive load data collecting devices,such as smart meters,

that deployed at the entrance of residents user power, sampling

the user’s total power load information (low frequency or high

frequency), and transmitted to the data analysis center for load

identiﬁcation through the communication lines. Hence, user

types of electrical equipment and start-stop information, such

as time, can be identiﬁed and used for the optimal dispatching

or precise load control of the power grid. At the same time, it

2711

The 6th IEEE Conference on Energy Internet and Energy System Integration

Authorized licensed use limited to: ZTE CORPORATION. Downloaded on October 26,2023 at 06:43:15 UTC from IEEE Xplore. Restrictions apply.

can also further obtain the characteristics of users’ electricity

consumption behavior and types of users, and provide users

with user portraits and better electricity experience.

Most Existing NILM methods can be divided into event

based NILM methods that focus on temporal data and non-

event based NILM methods that focus on steady data [7].

The event based NILM methods generally including event

detection and classiﬁcation steps [8], hence, the accuracy of

event detection inﬂuences the performance of NILM methods

directly. Three of the typically event detection methods are

CSD (cumulative sum detector) [9], GLRD ( generalized like-

lihood ratio detector) [10] and GFD (goodness-of-ﬁt detector)

[11], all of which require ﬁxed window size for reference

set and current set [8]. The non-event based NILM methods

usually take every sample of aggregated power data into

account [8], hence, the input data may be too large to analysis

and leading to computationally expensive and time consuming.

Although more progress has been made in the research of

non-intrusive load monitoring algorithms, good identiﬁcation

results have been achieved on some datasets. However, there

are few studies on data collection of non-intrusive load data,

and the current application of non-intrusive load monitoring

still has several shortcomings in data collection:

1) Large Data volume. Non-intrusive load data is often

months or even years of data. If we want to obtain better

load identiﬁcation effect, we need high frequency load

data. The continuously collected data makes the data

volume of non-intrusive load very large, which makes

the storage and computing resources required for non-

intrusive load identiﬁcation very high, which hinders the

application of non-intrusive load identiﬁcation.

2) Varies data length. The existing non-intrusive load data

collection method is a simple ﬁxed frequency, random

start/stop way to collect data. The collected data is a

single time series of varying lengths. Moreover, due to

the random and unﬁxed time when users use electrical

appliances, and a large amount of time (such as night)

when electrical appliances are not used, load ﬂuctuations

are randomly, discrete and non-uniform scattered in the

time series. This results in the variable length of the

collected non-intrusive load sample data, which affects

the performance of the non-intrusive load identiﬁcation

algorithm.

Hence, in this paper, a non-intrusive load data collec-

tion strategy based on event trigger and sliding-time-window

is proposed, by combining event-trigger and sliding-time-

window, the data volume of non-intrusive load collected can be

signiﬁcantly reduced while the size of data samples collected

each time is ﬁxed. The contribution of this paper can be

summarized as follows:

1) Through the event-triggered collection strategy, only the

data segment with load hopping is collected, which

greatly reduces the amount of non-intrusive load data

collected.

2) The data collecting method of sliding window ensures

that the collected data is a data segment of ﬁxed length,

and contains the complete data at the moment of load

change, which is more convenient for data processing.

3) The comparative experiment proves the effectiveness

of the method. On the REDD dataset [12], seven

non-intrusive load monitoring algorithms included in

NILMTK [13], [14] are used to compare the perfor-

mance of the data before and after using the acquisition

strategy. The experimental results show that the pro-

posed data collection strategy not only greatly reduces

the data size, but also improves the performance of the

non-intrusive load identiﬁcation algorithm.

In the following content of this paper, section II gives a

detailed analysis of the problems solved in this paper,and

introduces the framework of non-intrusive load data collection

strategy based on event-trigger and sliding-time-window with

its detailed data collecting process. section III introduces the

performance evaluation of the proposed algorithm on REDD

datasets, and finally concludes the paper in section IV.

II. PROPOSED DATA COLLECTION STRATEGY

A. Problem Introduction

The existing non-intrusive load data collection is mostly s-

ingle time series data, which can be divided into total load data

(meter data) and load data of single appliance. The acquisition

frequency varies from 1min to 250kHz, the collected data

include V(Voltage), I(Current), P(active-power), Q(Reactive

Power), S(passive-power), F (Frequency), PF (Power Factor),

One or more of E(Energy). The collection time varies from

several months to several years for low-frequency data, while

for high-frequency data, due to the large amount of data, the

published data is often only months or even weeks.

Research shows that high frequency data has great advan-

tages over low frequency data. For example, two devices of the

same type can even be distinguished using raw data sampled

at 1MHz. However, the acquisition of high frequency data is

accompanied by a thousand fold increase in the size of the

dataset. For example, the original high frequency data of uk-

dale (16kHz) generates 8.3GB of data per day, which results

in some high frequency data sets with only a few minutes

of data to control the size. This causes a lot of trouble for

researchers to study high-frequency NILM data. In terms of

the format of data sets, low-frequency data can often provide

original data, while high-frequency data usually only provide

compressed data (FLAC or WAV format) because the original

data is too large.

The collection frequency, duration and volume of data sets

commonly used in NILM are shown in the following Tab.I:

Therefore, if it can effectively reduce the size of NILM

dataset, it means that the terminal NILM data collection

terminal can collect and store data for longer time and more

appliances, while the cloud NILM data storage and analysis

will become easier and cheaper. At the same time, if the

NILM data can be effectively divided into ﬁxed size samples,

the input requirements of different algorithms can be better

adapted, and more algorithms can be better used for analysis.

2712