Cross-Modal Adaptation of Large Language Models for Building Energy Forecasting

Energy forecasting research has traditionally focused on individual buildings or clusters of similar facilities with limited progress on generalized models applicable across diverse building types. However, growing demands and greater variability in energy use patterns call for scalable and efficient solutions to enhance energy management. Barriers to more scalable solutions lie in the limited availability of large-scale, high-quality datasets, the heterogeneity of data attributes across different sources, and modeling efficiency. To this end, this study investigates the application of transfer learning through the cross-modal adaptation of pretrained large language models (LLMs), specifically GPT-2, for building energy forecasting. By fine-tuning GPT-2, the model is adapted to handle multivariate time series data and predict energy consumption across a dataset of 100 commercial buildings. The impact of input sequence length on forecasting performance was also evaluated, with results indicating that a 168-h (1-week) input length yields better performance. Moreover, comparative analyses reveal that the GPT-2-based approach achieves performance comparable to or exceeding that of state-of-the-art energy forecasting models while maintaining computational efficiency. These findings highlight the potential of cross-modality adaptations of pretrained large models, such as GPT-2, in advancing predictive energy management for the built environment as scalable and adaptable solutions.