Automated data-driven building energy load prediction method based on generative pre-trained transformers (GPT)

Generative pre-trained transformers (GPT) have shown remarkable capabilities in automated code generation for data-driven building energy load prediction scenarios, leading to substantial savings in time and costs. However, it is quite difficult for inexperienced users to provide high-quality prompts to GPT for generating satisfactory codes. To address this challenge, a GPT-based automated data-driven building energy load forecasting method is proposed in this study. Prompting functions are designed to automatically generate prompts for model training and deployment. Bayesian optimization is utilized to optimize the prompting functions for improving the prediction accuracy of GPT-generated codes. External knowledge bases are developed to improve the code correctness of GPT by adding additional knowledge to prompts. Furthermore, a self-correction strategy is designed to enable GPT to automatically correct errors in GPT-generated codes. This method is employed to forecast the energy loads of two real buildings for performance evaluation. GPT-3.5 is utilized in the evaluation process. The codes generated by this method exhibit high prediction accuracy, achieving an average R² of 0.95 for the two buildings. The code correctness of GPT-3.5 is increased by an average of 90.0 % by using the external knowledge bases. Moreover, the self-correction strategy effectively corrects some unpredictable mistakes made by GPT-3.5.