Value of LLM-Based Methods for Heat-Pump Fault Detection

Establish whether large language model-based approaches add value over established heat-pump fault-detection and diagnosis methods, using a dedicated benchmark that treats defrost-cycle labelling, part-load coefficient-of-performance trajectories, and refrigerant-state channels as first-class inputs.

Background

The review identifies a substantial application gap: none of the included studies addresses heat-pump fault detection, forecasting, or control using LLMs. Existing LLM-based fault-diagnosis knowledge bases and fine-tuning datasets primarily concern ASHRAE RP-1312 chiller and air-handling-unit benchmarks, whose fault signatures and sensor variables differ from those of heat pumps.

Heat-pump systems exhibit distinct defrost transients, variable-speed part-load behaviour, and reverse-cycle refrigerant-side faults. The review therefore cautions that transferring cooling-oriented fault libraries to heat pumps could produce plausible but physically incorrect explanations. Resolving the open question would require a benchmark incorporating heat-pump-specific operational and refrigerant-state information and comparing LLM-based systems against established heat-pump fault-detection methods.

References

Whether LLM-based approaches add value over established heat-pump FDD methods is therefore an open question.

— Large Language Models for HVAC Operations in Building Energy Systems: A Critical Review of Methods, Applications, and Deployment Readiness  (2609.05314 - Neubauer et al., 4 Sep 2026) in Section A1, “Fault Detection and Diagnosis (A1),” subsection “Scope and Corpus”