The development of deep learning in building energy: A bibliometric analysis

Deep learning is increasingly used in operational building energy applications, but the field’s interdisciplinary knowledge structure and its translation from algorithmic development to engineering practice remain insufficiently understood. This study analyzes 729 Web of Science Core Collection publications from 2015 to 2025 using Bibliometrix, VOSviewer, and CiteSpace. Publication growth, collaboration networks, journal citation flows, and thematic evolution were examined. Output accelerated after 2021, with 602 papers (82.6%) published during 2022–2025. China and the United States dominated collaboration, while Energy and Buildings, Applied Energy, and Journal of Building Engineering were the core publication outlets. Within the corpus, research attention shifted from energy and load prediction toward HVAC control, deep reinforcement learning, model predictive control, transfer learning, fault diagnosis, and digital-twin-enabled energy management. Citation flows revealed a three-layer knowledge-transfer pathway: computer science supplies computational methods, energy and environmental sciences frame energy problems, and building science translates them into applications. Beyond conventional descriptive mapping, this study connects this interdisciplinary pathway with the capabilities and deployment constraints of major model families, clarifying how computational methods are translated into building-energy applications. The findings identify transferability, physical consistency, interpretability, control safety, and long-term real-building validation as priorities for future research.