Leveraging Large Language Model for Generalization in Building Energy Management

Buildings are major contributors to global electricity consumption and emission, making the development of effective Building Energy Management Systems (BEMS) essential for improving energy efficiency, ensuring thermal comfort. Artificial Intelligence (AI)-based methods, particularly Deep Reinforcement Learning (DRL), have recently garnered significant attention in the BEMS. However, AI-based methods often face challenges related to generalization across diverse building environments. This paper introduces a novel approach for BEMS control, namely Sequence-Based Building Energy Modeling Framework (SeqBECMF), which models the energy management problem as a sequence modeling task. SeqBECMF leverages the power of large language model (LLM) to model long-term dependencies in building energy data. To further enhance adaptability across varying building environments, based on SeqBECMF, we propose the Building Adaptation Decision Transformer (BADT), which incorporates a building adapter module to scale actions based on specific environmental needs. BADT uses a generative pre-trained transformer (GPT) backbone to improve temporal modeling and represent diverse building conditions effectively. We evaluate the performance of the BADT framework on the CityLearn dataset and compare it to state-of-the-art DRL algorithms. Experimental results show that BADT consistently outperforms these DRL methods by minimizing key adverse metrics, including thermal discomfort hours, grid ramping, and electricity consumption. Moreover, BADT demonstrates strong generalization to unseen buildings, setting a new standard for scalable and adaptable BEMS solutions.