Abstract
The operation of modern buildings generates large volumes of sensor data that, if effectively utilized, can significantly improve energy efficiency and fault detection. However, conventional fault detection and diagnosis (FDD) systems often rely on rule-based methods that lack context-awareness, generate excessive false alarms, and demand expert interpretation. This paper presents a novel framework that combines Large Language Models (LLMs), time series data, and semantic knowledge graphs to enhance building operation and diagnostics. The approach is implemented and evaluated in the ZEB Laboratory, a full-scale zero-emission building and research infrastructure in Trondheim, Norway.
The proposed system integrates structured sensor metadata from a Neo4j-based knowledge graph with real-time measurement data from an InfluxDB time-series database. An LLM (GPT-4o) processes this data to detect anomalies and generate contextual explanations and recommended actions in natural language. One exemplary room serves as the primary testbed, showcasing how knowledge-driven reasoning enables accurate, actionable alerts with reduced false positives. Test scenarios demonstrate the framework’s ability to identify faults and guide operational decisions. Structured outputs support seamless dashboard integration and potential for automated control. This work shows how generative AI combined with semantic metadata and real-time data streams can transform building management into a more intelligent, interpretable, and proactive process.