Investigating the application of a commercial and residential energy consumption prediction model for urban Planning scenarios with Machine Learning and Shapley Additive explanation methods

Abstract

Building energy forecasting methodologies utilized by municipal governments tend to be geared heavily towards depicting broader qualitative representations of regional change and are in need of complementary data-driven models that can produce quantitatively reliable depictions of future energy consumption at the neighborhood-level. The current research demonstrates an application of a Machine Learning (ML) model in the form of an Extreme Gradient Boosting (XGBoost) algorithm for forecasting the energy use of commercial and residential buildings. The methodology serves to improve on municipal scenario planning by providing a more spatially granular representation of future energy use. In this way, city government and urban planners can more accurately set carbon emission benchmarks and target specific locales for sustainability initiatives. The second major contribution of the study is to demonstrate how scenario planning approaches can utilize existing Machine Learning techniques to compensate for gaps in the data. This work is developed through a case study of Philadelphia. The study begins with the construction of residential and commercial energy models for the year 2015 and corresponding models for the year 2045. The forecast models integrate regional socioeconomic trends from the Delaware Valley Regional Planning Commission (DVRPC) scenario of Enduring Urbanism. The commercial energy use model is developed from the DVRPC’s open-source Geographic Information System (GIS) datasets, the Commercial Buildings Energy Consumption Survey (CBECS), and CoStar commercial real estate data. The residential model applies the Residential Energy Consumption Survey (RECS), the Public Use Microdata Sample (PUMS), and Census Bureau American Community Survey (ACS) estimates. A corresponding SHAP (SHapley Additive exPlanations) analysis is implemented to pinpoint feature contributions to the model’s energy estimates. By using the PopGen software, the model’s energy estimates could be analyzed at the household level, the smallest possible scale. To provide a useful resource for key stakeholders, the study aggregates model output by Traffic Analysis Zone (TAZ) and Public Use Microdata Area (PUMA) to display a detailed forecast of energy use. The results indicate that the DVRPC Enduring Urbanism trends in income and employment do not significantly affect energy consumption for the study area. However, features related to lower building intensity (e.g., lower square footage, fewer floors per building) were associated with reduced energy use in both models. Additionally, the study found residential buildings under the “single-family attached” zoning designation to correspond with higher energy estimates.