Transfer Learning-Supported Building Electric Energy Forecasting Using Neural Network Models
Published:
The operation data recorded in the Building Automation System (BAS) provides an untapped opportunity for developing data-driven models to improve building performance. However, BAS data might be insufficient to derive an accurate model for practical applications. For example, the data quality may be unreliable, and for new buildings or renovations, operation data may simply be unavailable. To tackle this problem, this paper investigates transfer learning to forecast cooling electricity demand and total building electricity demand using neural network models. The case study includes two buildings, a source building with synthetic data from an EnergyPlus model and a target building with real-world BAS data. A neural network model is first pre-trained with the synthetic data from the source building, and then transferred to the target building to forecast the electricity demand of the cooling system and the whole building. Compared with the model derived only from the BAS data of the target building, the transfer learning approach shows an improved forecasting performance by 40% and 44% respectively for the cooling and total building electricity demand. The transfer learning results could be further used in advanced controls such as model predictive control.
Recommended citation: Dou, H., Zhang, K. (2024). Transfer Learning-Supported Building Electric Energy Forecasting Using Neural Network Models. In: Proceedings of ASim Conference 2024: 5th Asia Conference of IBPSA, December 8–10, 2024, Osaka, Japan.