Building Energy Performance Prediction Using Machine Learning: A Data-Driven Decision-Making Framework for Energy Retrofits: Fid-Bau Portal

Building Energy Performance Prediction Using Machine Learning: A Data-Driven Decision-Making Framework for Energy Retrofits

Murrieum, Munahil / Jafari, Amirhosein / Akhavian, Reza

Abstract

Buildings are major consumers of energy in the United States and worldwide. The first step in building energy efficiency improvement is to predict the consumption. Data-driven models provide a practical approach to energy consumption prediction; however, the energy data often have numerous variables creating a high-dimension feature space which poses challenges for data analysis. The primary goals of this study are to present a framework that (1) evaluates and selects the most relevant features regarding building energy performance, and (2) develops an energy consumption predictive model using the selected features. To illustrate the application of the proposed model, it is implemented on the U.S. Energy Information Administration (EIA) dataset of 2015 Residential Energy Consumption Survey (RECS), which is a nationally representative sample of housing units. First, a combination of statistical analysis and expert judgment is used to develop a model to determine the building energy-related features with the highest impact on building energy performance, based on the pre-defined energy consumption categories (e.g., space heating, lighting, air conditioning, etc.). Then, different machine learning methods are trained to develop an accurate energy consumption prediction model that relates the selected building energy-related features to the building energy consumption, based on different energy consumption categories. In terms of explanatory power, the R-square value of the prediction model is 67.0% which supports overall model validity.