Mar 04

Deliverable 5.1: Statistical learning of residential electricity consumption data

Residential and commercial actors contribute around 30% of global energy-related emissions. By transitioning from consumers to prosumagers, the energy consumption and GHG emissions of these two sectors are expected to be reduced. This transition can be supported, among other factors, by the digitalization of energy consumption and production systems within buildings. Smart Meters (SMs) are one of the main technical innovations supporting digital information collection and analysis for optimizing energy management.

The availability of big datasets from electricity smart meters in the residential sector not only enables novel types of behavioral and other demand response interventions, but also provides an unprecedented opportunity for a new empirical basis for energy-demand policy making. The study of high-frequency electricity consumption data can help understand patterns of consumption at weekly, daily and hourly level, where the most common mean of understanding household behavior has been derived from monthly billing.

Given the large size of SM datasets, machine-learning tools become indispensable to process the collected data points and distill useful information to be fed in the modeling and policy evaluation pipeline. One of these tools is clustering, which is the main focus of this report. Clustering enables the extraction of relevant patterns and behaviors in a concise manner, offering the opportunity to describe an otherwise large ensemble of data points in terms of a small number of representative statistical variables.

The report reviews the relevant literature addressing the problem of clustering high-frequency electricity consumption data. All the technical aspects of clustering are taken into account and critically discussed to better understand, design and use clustering analyses. Then, three case studies are presented. These leverage on recently collected smart meter datasets in two countries, Italy and Poland, and showcase potential alternative choices in designing a clustering strategy.

This is the first study aimed at using load profile clustering and socio-demographic segmentation to improve energy demand modeling at EU level, in particular by better capturing energy consumption behaviors, comparing different clustering approaches, and including very recent high-frequency electricity consumption data.

Full report available here.