Artificial intelligence for an enhanced as-is BIM energy analysis: Enabling an efficient energy retrofit through the automation of the scan-to-BIM process

In the last few years, energy retrofit of the existing building stock has quickly positioned itself centre stage of research and application, with several European directives and countries’ legislations focusing the topic in pursuit of sustainability and economic goals. Similarly, as a prominent topic within the Architecture, Engineering and Construction industry, Building Information Modelling (BIM) has risen in importance in scientific and practical communities, morphing itself from a futuristic scholar concept into a vital core piece of the industry. With an increasing overlap of both areas, researchers and practitioners focused the As-Is BIM Energy Analysis (AIBEA) of a building as an answer to the slow progress of retrofitting the existing building stock, enabling the quick analysis of a building’s energy consumption and the exhaustive comparison of constructive solutions to increase its efficiency. However, several obstacles still hinder the overall productivity and accuracy of this process, with multiple scientific works denoting a growing need to swiftly and automatically acquire accurate, structured, and semantically enriched three-dimensional digital models of existing buildings. To this end, the present thesis aims to contribute with an answer to this problem by tackling it from two perspectives: (1) the development of supporting documentation and tools for an accurate and swift AIBEA, and (2) the integration and automation of the scan-to-BIM process. To achieve this goal, the author relies on two primary technologies: laser scanning and Artificial Intelligence (AI). These are applied to accurately acquire the as-is building geometry and automate the resulting point cloud segmentation, classification, and modelling within a BIM authoring environment. Initially, this thesis starts by reviewing the state of the art and theoretical basis on AIBEA and scan-toBIM, exploring multiple related topics as existing legislation and supporting documentation for the AIBEA workflow; as-is building geometric and energy-related data acquisition; existing software tools and interoperability; AI; deep learning approaches to point cloud segmentation and classification; as well as automated BIM modelling and subsequent data enrichment. Afterwards, the thesis’ contributions are presented and divided into five unique modules, allowing for a thorough exploration of each contribution. The modules are: (1) identification of AIBEA requirements; (2) as-is building data acquisition; (3) point cloud segmentation and classification; (4) automated BIM modelling; and (5) BIM model enrichment and exportation to energy analysis software. Together, the modules comprise a methodology for an enhanced AIBEA. This methodology is then applied in five different experiments to evaluate its performance and identify its advantages and limitations. Based on the achieved results, relevant conclusions regarding the thesis’ contributions and applied technologies are retrieved. The experiments achieved successful results, justifying its continuous development in future works. Throughout the thesis, multiple topics of further interest to the literature are expanded, promoting the research of existing scientific and industry problems, and proposing original methods for the identification of contractual requirements and quality verification parameters for the scan-to-BIM process; analysis of laser scanner parameters and its influence over the final point cloud information; optimal placement of laser scanner stations; artificial training of deep learning algorithms; and identification of construction materials through laser scanning.