Large Language Models as Tools for Public Building Energy Management: An Assessment of Possibilities and Barriers

Abstract

This study examins usability of large language model-based (LLM) chat bots, specifically GPT-3.5 and GPT- 4, as assisting tools in the energy manageemnt of educational buildings in Serbia and Poland. The assessment is based on the comparison of three key usability aspects: 1) the accuracy of expert opinion replication in classifying building construction periods, 2) familiarity with field-specific legislation; and 3) knowledge of details regarding building thermal envelopes. In replicating constuction periods and sugesting legislation, LLM chat bots performed admirably in Poland, but less so in Serbia. Regarding thermal enveople characteristics, GPT-3.5 indicated U-value spans encompassing actual values, but to broad to be useful. U-value spans provided by GPT-4 were narrower, but they did not generally intersect with the actual U-value range. The study concludes that LLM chat bots indicate great potential for assisting non-experts in public building energy management, with usability varying by country, but still far from experts.