A Comparative Analysis of Fine-Tuning and Prompt Engineering for LLM-Based Automation of Building Energy Models for Rooftop Photovoltaics

Department of Convergence and Fusion System Engineering, Kyungpook National University, South Korea

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

This study examines the effectiveness of Large Language Models (LLMs) in automating the generation of EnergyPlus Input Data Files (IDFs) for buildings equipped with integrated rooftop photovoltaic (PV) systems. While LLMs offer a promising future for Building Energy Modeling (BEM), their ability to handle the strict geometric constraints of energy systems remains underexplored. This study focuses on rooftop PV configurations, including tilt angles, inter-row spacing (gap), and building size. The performance of parameter-efficient fine-tuning (LoRA) and standard prompt engineering (zero-shot, one-shot, and two-shot) is evaluated across three transformer models: GPT-2 Medium, Facebook OPT-350M, and Flan-T5 Base. The results demonstrate that standard prompt engineering fails to produce executable simulations for complex PV-integrated models. In contrast, the fine-tuned models successfully design for both building and PV configurations. Furthermore, we validated the practical utility of the fine-tuned models by conducting a design-optimization case study in San Francisco and Chicago, successfully identifying optimal PV tilt angles and spacing configurations. This research confirms that fine-tuning small-scale LLMs is a viable and robust strategy for automating complex BEM tasks.