Transfer learning and parameter-efficient fine-tuning for heating energy consumption prediction using urban building energy models (UBEM)

To estimate the energy demand of buildings on a neighborhood scale, physics-based Urban Building Energy Modeling (UBEM) has emerged as a reliable method. UBEM simulations can also be used to develop training datasets for machine learning-based approaches that can predict energy performance. However, collecting large training datasets and developing UBEMs for each district is too resource-intensive and, therefore, not practical. Transfer learning (TL) can address this limitation by transferring knowledge from an existing model already trained with a large dataset to train another target model in another district with a smaller dataset size. In this paper, we present an approach for TL-based prediction of building heating energy consumption on an urban scale based on four existing UBEMs of cities with different climatic conditions and building characteristics. To this end, in addition to the conventional approach of fine-tuning an ML model for a new city, we evaluate the potential impact of parameter-efficient approaches (namely, LoRA and LoRA+) as they are expected to provide better TL with less data. Our experiments show that TL does provide better estimation compared to training ML models from scratch. Furthermore, our results suggest that this gain in accuracy is even higher if the target city has a substantially low number of buildings to train an ML model from scratch. Finally, our experiments provide evidence for better TL with parameter-efficient approaches.