Exploring the integration of simulation and deep learning models for urban building energy modelling and retrofit analysis

Abstract

As the world rapidly urbanizes, cities will need to be created and expanded to accommodate their growing populations - putting immense pressure on engineers, scientists and policymakers to improve the efficiency of the energy intensive built environment. One of the key barriers to improving the energy efficiency of cities is the ability to accurately model and characterize the energy performance of their buildings. While simulation-based methods have been developed to help predict the energy consumption of urban buildings, they are limited in their ability to quickly evaluate the effects of various design or retrofit scenarios. New data-driven methods are emerging to model building energy usage, but their lack of an underlying physics-based engine limits applicability and interpretability for assessing design or retrofit scenarios. In this paper, we employ the use of an integrated simulation and data-driven method (i.e., Data-driven Urban Energy Simulation or DUE-S) to model a large-scale retrofit policy on a case study of 52 buildings in a Californian city. Our results indicate that the DUE-S model is able to capture the energy impacts that the urban context has on buildings that undergo retrofits as well as those that do not. Our primary contribution is to demonstrate the merits of combining physics-based building simulation methods with new data-driven machine learning methods (i.e., transfer learning) to assess the impact of various design and retrofit scenarios across a large urban area and in turn spawn future research at the intersection of simulation and data science. In the end, realizing deep energy savings from urban buildings will require new tools that are both accurate and interpretable enough to inform decision-making for a variety of urban sustainability stakeholders regarding early stage designs, energy efficiency retrofits and environmental policymaking.