Generative pre-trained transformers (GPT)-based automated data mining for building energy management: Advantages, limitations and the future

Abstract

Advanced data mining methods have shown a promising capacity in building energy management. However, in the past decade, such methods are rarely applied in practice, since they highly rely on users to customize solutions according to the characteristics of target building energy systems. Hence, the major barrier is that the practical applications of such methods remain laborious. It is necessary to enable computers to have the human-like ability to solve data mining tasks. Generative pre-trained transformers (GPT) might be capable of addressing this issue, as some GPT models such as GPT-3.5 and GPT-4 have shown powerful abilities on interaction with humans, code generation, and inference with common sense and domain knowledge. This study explores the potential of the most advanced GPT model (GPT-4) in three data mining scenarios of building energy management, i.e., energy load prediction, fault diagnosis, and anomaly detection. A performance evaluation framework is proposed to verify the capabilities of GPT-4 on generating energy load prediction codes, diagnosing device faults, and detecting abnormal system operation patterns. It is demonstrated that GPT-4 can automatically solve most of the data mining tasks in this domain, which overcomes the barrier of practical applications of data mining methods in this domain. In the exploration of GPT-4, its advantages and limitations are also discussed comprehensively for revealing future research directions in this domain.