Multi-modal prediction of building energy efficiency using apartment listing data

Predicting the energy efficiency of buildings at scale when detailed building data is limited remains a challenge. Tools for doing so can enhance our understanding of the building stock, which can be used for decision-making and policymaking related to building renovation for efficiency. Such tools, when applied in the residential con-text, could also enhance access to energy-related information for home renters and buyers when making decisions on where to live. In this work, we build a multi-modal machine learning model to predict the energy efficiency of residential units. We leverage a large dataset of tens of thousands of buildings in Switzerland that include tabular, text, and image data that can commonly be found in online apart-ment listing sites. After regressing out weather, occupancy, and energy price in-formation, our model predicts the inherent energy efficiency fixed effects with about a 30% normalised root mean squared error. We explore the relative im-portance of each data modality and demonstrate how the multiple modalities can be integrated into an overall ensemble approach. This research shows how more information about inherent building energy efficiency can be disclosed through ma-chine learning tools.