Data-driven residential building energy consumption prediction for supporting multiscale sustainability assessment

Sustainability assessment has drawn much attention worldwide. However, current sustainability assessment models/systems are usually focused on a single scale, which does not allow for capturing the interdependencies across the different scales. Also, existing sustainability assessment models/systems typically require high levels of data collection and complex measurements, which is usually costly and time-consuming. To address these gaps, the authors propose a data-driven multiscale sustainability assessment and prediction (DM-SAP) framework, which (1) focuses on the building, neighborhood, and city scales, and (2) takes a data-driven approach that relies on machine learning. The machine learning approach focuses on two primary objectives: (1) learning from history (previous building, neighborhood, and city data) to predict and assess sustainability metrics (e.g., energy consumption) in an efficient and reliable manner, and (2) feature selection to identify the minimum amount of data that would be sufficient for reliable assessment. This paper focuses on presenting the feature selection and the machine learning-based model development for predicting residential building energy consumption for supporting multiscale sustainability assessment. The LASSO algorithm was used for feature selection. Three machine learning algorithms were implemented and tested. The prediction performance was evaluated in terms of coefficient of variation. The 2009 Residential Energy Consumption Survey (RECS) by the U.S. Energy Information Administration (EIA) was utilized for model training and testing. The testing results showed reasonable prediction performance.