PNNL

Abstract

Mapping BACnet point names to standardized Brick classes is essential for interoperability and analytics in building automation systems, yet manual tagging remains time-consuming and error prone. Automating this process can reduce labor costs, improve scalability across large building portfolios, and enable advanced analytics for energy efficiency and diagnostics. Previous approaches have relied on classifiers trained on curated datasets such as Mortar or on rule-based methods, both of which struggle with rare abbreviations, vendor-specific conventions, and heterogeneous building contexts. In this work, we propose an automated semantic tagging framework that integrates classifier-based normalization, embedding-driven semantic matching, and foundational language models, extended with retrieval-augmented prompting (RAP) to inject domain-specific context. We evaluate the framework across eight anonymized PNNL-operated buildings comprising more than 7,429-point names, with classifiers trained on Mortar and predictions validated against SME-curated ground truth. Results show that classifiers achieved 28.3% accuracy, foundational models (Llama3) achieved 26.6%, and RAP improved accuracy to 35.6%, confirming that contextual grounding is critical to robust inference. Building on these results, future work will explore contextual enrichment using metadata and time-series signals, expansion of training corpora, transfer learning for domain adaptation, and flexible top-k recommendations to further reduce reliance on manual tagging.