Data-driven modelling of building retrofitting with incomplete physics: A generative design and machine learning approach

Abstract

Building performance simulation (BPS) based on physical models is a popular method for estimating the expected energy savings from energy-efficient building retrofitting. However, for many buildings, especially older buildings, built several decades ago, an operator do not have full access to the complete information for the BPS method. Incomplete information comes from the lack of detailed building physics, such as the thermal transmittance of some building components due to the deterioration of components over time. To address this challenge, this paper proposed a data-driven approach to support the decision-making of building retrofitting selections under incomplete information conditions. The data-driven approach integrates the backpropagation neural networks (BRBNN), fuzzy C-means clustering (FCM), and generative design (GD). It generates the required big database of building performance through generative design, which can overcome the problem of incomplete information during building performance simulation and energy-efficient retrofitting. The case study is based on old residential buildings in severe cold regions of China, using the proposed approach to predict energy-efficient retrofitting performance. The results indicated that the proposed approach can model the performance of residential buildings with more than 90% confidence, and show the variation of results. The core contribution of the proposed approach is to provide a way of performance prediction of individual buildings resulting from different retrofitting measures under the incomplete physics condition.