Multi-resolution load profile clustering for smart metering data

Ran Li, Furong Li, Nathan D. Smith

Research output: Contribution to journalArticle

20 Citations (Scopus)

Abstract

This paper proposes a novel multi-resolution clustering (MRC) method that for the first time classifies end customers directly from massive, volatile and uncertain smart metering data. It will firstly extract spectral features of load profiles by multi-resolution analysis (MRA), and then cluster and classify these features in the spectral-domain instead of time-domain. The key advantage is that the proposed method will allow a dynamic load profiling to be flexibly re-constructed from each spectral level. MRC addresses three key limitations in time-series based load profiling: i) large sample size: sample size is reduced by a novel two-stage clustering, which firstly clusters each customers' massive daily profiles into several Gaussian mixture models (GMMs) and then clusters all GMMs; ii) volatility: it avoids the interferences between different load features (e.g. magnitude, overall trend, spikes) by decomposing them onto different resolution levels, and then clustering separately; iii) uncertainties: as the GMM can give a probabilistic cluster membership instead of a deterministic one, an additive classification model based on the posterior probability is proposed to reflect the uncertainty between days. The proposed method is implemented on 6369 smart metered customers from Ireland, and compared with the load profiles used by the U.K. industry and traditional K-means clustering. The results show that the developed MRC outperformed the traditional methods in its ability in profiling load for big, volatile and uncertain smart metering data.
Original languageEnglish
Article number TPWRS-00618-2015
Pages (from-to)4473-4482
Number of pages10
JournalIEEE Transactions on Power Systems
Volume31
Issue number6
Early online date8 Mar 2016
DOIs
Publication statusPublished - 1 Nov 2016

Fingerprint

Multiresolution analysis
Dynamic loads
Time series
Industry
Uncertainty

Cite this

Multi-resolution load profile clustering for smart metering data. / Li, Ran; Li, Furong; Smith, Nathan D.

In: IEEE Transactions on Power Systems, Vol. 31, No. 6, TPWRS-00618-2015, 01.11.2016, p. 4473-4482.

Research output: Contribution to journalArticle

@article{23f5eca4c9104cabb1fd4682532f1bba,
title = "Multi-resolution load profile clustering for smart metering data",
abstract = "This paper proposes a novel multi-resolution clustering (MRC) method that for the first time classifies end customers directly from massive, volatile and uncertain smart metering data. It will firstly extract spectral features of load profiles by multi-resolution analysis (MRA), and then cluster and classify these features in the spectral-domain instead of time-domain. The key advantage is that the proposed method will allow a dynamic load profiling to be flexibly re-constructed from each spectral level. MRC addresses three key limitations in time-series based load profiling: i) large sample size: sample size is reduced by a novel two-stage clustering, which firstly clusters each customers' massive daily profiles into several Gaussian mixture models (GMMs) and then clusters all GMMs; ii) volatility: it avoids the interferences between different load features (e.g. magnitude, overall trend, spikes) by decomposing them onto different resolution levels, and then clustering separately; iii) uncertainties: as the GMM can give a probabilistic cluster membership instead of a deterministic one, an additive classification model based on the posterior probability is proposed to reflect the uncertainty between days. The proposed method is implemented on 6369 smart metered customers from Ireland, and compared with the load profiles used by the U.K. industry and traditional K-means clustering. The results show that the developed MRC outperformed the traditional methods in its ability in profiling load for big, volatile and uncertain smart metering data.",
author = "Ran Li and Furong Li and Smith, {Nathan D.}",
year = "2016",
month = "11",
day = "1",
doi = "10.1109/TPWRS.2016.2536781",
language = "English",
volume = "31",
pages = "4473--4482",
journal = "IEEE Transactions on Power Systems",
issn = "0885-8950",
publisher = "IEEE",
number = "6",

}

TY - JOUR

T1 - Multi-resolution load profile clustering for smart metering data

AU - Li, Ran

AU - Li, Furong

AU - Smith, Nathan D.

PY - 2016/11/1

Y1 - 2016/11/1

N2 - This paper proposes a novel multi-resolution clustering (MRC) method that for the first time classifies end customers directly from massive, volatile and uncertain smart metering data. It will firstly extract spectral features of load profiles by multi-resolution analysis (MRA), and then cluster and classify these features in the spectral-domain instead of time-domain. The key advantage is that the proposed method will allow a dynamic load profiling to be flexibly re-constructed from each spectral level. MRC addresses three key limitations in time-series based load profiling: i) large sample size: sample size is reduced by a novel two-stage clustering, which firstly clusters each customers' massive daily profiles into several Gaussian mixture models (GMMs) and then clusters all GMMs; ii) volatility: it avoids the interferences between different load features (e.g. magnitude, overall trend, spikes) by decomposing them onto different resolution levels, and then clustering separately; iii) uncertainties: as the GMM can give a probabilistic cluster membership instead of a deterministic one, an additive classification model based on the posterior probability is proposed to reflect the uncertainty between days. The proposed method is implemented on 6369 smart metered customers from Ireland, and compared with the load profiles used by the U.K. industry and traditional K-means clustering. The results show that the developed MRC outperformed the traditional methods in its ability in profiling load for big, volatile and uncertain smart metering data.

AB - This paper proposes a novel multi-resolution clustering (MRC) method that for the first time classifies end customers directly from massive, volatile and uncertain smart metering data. It will firstly extract spectral features of load profiles by multi-resolution analysis (MRA), and then cluster and classify these features in the spectral-domain instead of time-domain. The key advantage is that the proposed method will allow a dynamic load profiling to be flexibly re-constructed from each spectral level. MRC addresses three key limitations in time-series based load profiling: i) large sample size: sample size is reduced by a novel two-stage clustering, which firstly clusters each customers' massive daily profiles into several Gaussian mixture models (GMMs) and then clusters all GMMs; ii) volatility: it avoids the interferences between different load features (e.g. magnitude, overall trend, spikes) by decomposing them onto different resolution levels, and then clustering separately; iii) uncertainties: as the GMM can give a probabilistic cluster membership instead of a deterministic one, an additive classification model based on the posterior probability is proposed to reflect the uncertainty between days. The proposed method is implemented on 6369 smart metered customers from Ireland, and compared with the load profiles used by the U.K. industry and traditional K-means clustering. The results show that the developed MRC outperformed the traditional methods in its ability in profiling load for big, volatile and uncertain smart metering data.

U2 - 10.1109/TPWRS.2016.2536781

DO - 10.1109/TPWRS.2016.2536781

M3 - Article

VL - 31

SP - 4473

EP - 4482

JO - IEEE Transactions on Power Systems

JF - IEEE Transactions on Power Systems

SN - 0885-8950

IS - 6

M1 - TPWRS-00618-2015

ER -