Electricity Consumption Clustering Using Smart Meter Data

Abstract

Electricity smart meter consumption data is enabling utilities to analyze consumption information at unprecedented granularity. Much focus has been directed towards consumption clustering for diversifying tariffs; through modern clustering methods, cluster analyses have been performed. However, the clusters developed exhibit a large variation with resulting shadow clusters, making it impossible to truly identify the individual clusters. Using clearly defined dwelling types, this paper will present methods to improve clustering by harvesting inherent structure from the smart meter data. This paper clusters domestic electricity consumption using smart meter data from the Danish city of Esbjerg. Methods from time series analysis and wavelets are applied to enable the K-Means clustering method to account for autocorrelation in data and thereby improve the clustering performance. The results show the importance of data knowledge and we identify sub-clusters of consumption within the dwelling types and enable K-Means to produce satisfactory clustering by accounting for a temporal component. Furthermore our study shows that careful preprocessing of the data to account for intrinsic structure enables better clustering performance by the K-Means method.

Multi-Domain Feature Extraction for Improved Clustering of Smart Meter Data

“…Four clustering techniques, i.e., random forest approach, k-Nearest Neighbour, decision tree and artificial neural network, are compared and it is found that random forest clustered the data better than others. But due to high computational cost and redundancy in data, instead of using raw data, extraction of salient features is considered an important step before clustering [2] [4][5][6][7][8][9][10] [12][13][14]. Features can be found in time [2] [7,8] [13], frequency [8] and time frequency [9] domains in order to have better knowledge about characteristics of meter data.…”

“…The cluster centres along with various distance functions (Canberra, manhattan, Euclidean and pearson) were used to estimate missing and future values by using other data in the same cluster. For harvesting the inherent structure from smart meter data, autocorrelation was applied with 24 lags on one week smart meter data in [12]. Then k means clustering technique was used on reduced dataset and it was found that for 32,241 smart meter data (over a period of one week) the optimal number of clusters were 12 (found by evaluating DBI mean index).…”

Statistical Approach in Analyzing of Advanced Metering Data in Power Distribution Grid

“…In [7] and [8] is explained benchmarking of smart meter data analytics. In [9] - [11] are presented clustering techniques for smart meter data analysis. Using clustering, customer loads with similar profiles can be identified.…”

“…[9]. In [11] is stated that is hard to identify individual clusters. Opposite this below, this paper presents one new statistical approach for smart meter data analysis.…”

“…Smart meter data analyzing in literature is so far usually used in load profile clustering [9][10][11]. Besides that, smart metering is used in distribution feeder modeling and distribution grid analysis improving.…”