Transfer learning for cross-building forecasting of building energy and indoor air temperature in model predictive control applications

When applying Model Predictive Control (MPC) for Heating, Ventilation and Air Conditioning (HVAC) systems in buildings, accurate forecasting of short-term energy demand and indoor air condition profiles is essential. However, new or retrofitted buildings lack sufficient operation data to develop precise data-driven models. This study investigates transfer learning techniques to enhance the forecasting performance of black-box models under limited data conditions. Specifically, we leverage synthetic data from an open-source EnergyPlus building model to pre-train three neural network models, which are then transferred to a real building and fine-tuned with limited measurements. The results indicate that incorporating synthetic data into the pre-training phase significantly enhances the forecasting accuracy for building and HVAC energy, as well as indoor air temperature profiles, over a 12-hour horizon with 15-minute intervals. The study underscores the potential of combining transfer learning with synthetic data to address data limitations, extending the applicability of learning-based MPC in real-world buildings. • Build unified models to forecast both building electrical load and thermal comfort • Improved forecasting accuracy in data-scarce target buildings using synthetic data • Compared three different transfer learning strategies and three model types • Transfer learning improved the performance of neural network models by 41%