Generation of whole building renovation scenarios using variational autoencoders

Abstract

Buildings consume a huge amount of energy, resulting in a considerable impact on the environment. In Canada, almost 70% of the total energy used by the commercial and institutional sectors was consumed by Heating, Ventilation and Air-Conditioning (HVAC) and lighting systems, which makes them the main targets of energy performance optimization methods. Furthermore, based on a governmental report, 40% of Quebec university buildings are in poor or very poor shape regarding structure and materials, and require immediate renovation. Therefore, it is of utmost importance to reduce energy consumption, and this can be accomplished by improving the design of new buildings or by renovating existing ones. Moreover, Simulation-Based Multi-Objective Optimization (SBMO) models can be used for optimizing and assessing different renovation scenarios considering Total Energy Consumption (TEC) and Life Cycle Cost (LCC). The time-consuming nature of SBMO has triggered the development of simplified and surrogate models within the design process. This study proposes a generative deep learning building energy model using Variational Autoencoders (VAEs), which could potentially overcome the current limitations. The proposed VAEs extract deep features from a whole building renovation dataset and generate renovation scenarios considering TEC and LCC of the existing institutional buildings. The proposed model also has the generalization ability due to its potential to reuse the dataset from a specific case in similar situations. The performance of the developed model has been demonstrated using a simulated renovation dataset to prove its potential. The results show that using generative VAEs is acceptable considering computational time and accuracy.