Discovery of Energy Performance Patterns for Residential Buildings Through Machine Learning

The building sector in New York City (NYC) accounts for about 75% of the greenhouse gas emissions (GHGE), and 40% of total energy consumption with an increasing rate of 5%. In a megacity like New York, tackling the lack of efficiency and sustainability in residential buildings requires identification, prediction, and analysis of energy performance patterns throughout the wide variation of building characteristics, location, and energy-related historical data. This study aims to discover and analyze energy performance patterns for residential buildings, and the method applied is a combination of supervised and unsupervised learning. The proposed method for the discovery of building energy performance patterns comprises three main variables: weather normalized (WN) site energy use intensity (EUI); GHGE Intensity; and energy efficiency grade. Four years of historical open data, from 2016 to 2019, was retrieved from various sources and combined into a dataset of 30.3 k data-points, covering 23 attributes. The developed models are verified against previous studies in terms of accuracy, achieving for site EUI and GHGE a reliable performance with an accuracy of 92%, and R² coefficients of about 0.85. The energy efficiency grade prediction model presented lower performance with nearly 80% accuracy. City planners, building designers, owners, and facility managers can benefit from the findings to track, manage, and improve building energy efficiencies through the implementation of renewable energies or other solutions, to achieve NYC’s Council goal of reducing GHGE by 80% by 2050, yet meeting the energy demands of the building infrastructure without relying exclusively on non-renewable resources.