Abstract
Buildings account for approximately 40% of global energy use, largely due to heating, ventilation, and air conditioning (HVAC) systems. Advanced control strategies, such as model predictive control (MPC), are crucial for optimizing energy usage in smart buildings. For effective MPC, high-fidelity models are necessary to reliably predict thermal responses under varying conditions. Low development costs, scalability and ability to process high-dimensional data and capture non-linear relationships make deep neural networks (DNNs) apt for predicting complex building dynamics. However, developing DNNs for new buildings is challenging due to their extensive data requirements. This paper investigates the use of transfer learning to address this issue, based on studies of different building models in different Norwegian climates. By pre-training control-oriented deep neural network models on synthetic building operation datasets, developed using EnergyPlus, from a similar source building in a similar climate, these models can be fine-tuned for new target buildings with minimal data. The study makes three key contributions. First, it introduces and evaluates six fine-tuning strategies for pre-trained DNNs, offering empirical insights into optimal approaches for adapting complex encoder-decoder architectures. Second, it employs two custom key performance indicators to quantify the effectiveness of transfer learning strategies, providing a standardized framework for assessing transfer learning in building dynamics prediction. Third, the study demonstrates the importance of fine-tuning specific model components, such as decoder layers, and the benefits of adding dense layers or gated recurrent units, especially in cold climates, where introducing colder weather data significantly improves predictive accuracy. Key findings include better results from fine-tuning decoder layers, especially early ones, and the importance of adding new layers before fine-tuning specific ones. Target buildings in mild climates and with similar physical properties (U-values) showed more stable and faster improvement from transfer learning. For the cold-climate case studies, introducing colder weather in the fine-tuning datasets significantly improved prediction accuracy.