Uncertainty quantification-based robust deep learning for building energy systems considering distribution shift problem

Artificial intelligence becomes one of the key technologies in building energy conservation, where deep learning algorithms provide feasible solution for the modeling problem of complex building energy systems. However, existing deep learning models are lack of robustness under distribution shift scenarios, which greatly limits their real-world application. To tackle this problem, first, we enrich the concept of distribution shift problem in building energy domain. And a cluster-based dataset splitting method is proposed to simulate the distribution shift scenarios. Second, we adopt uncertainty quantification methods to improve the overall robustness of deep learning model. Comprehensive data experiments are conducted based on the reference modeling problem of chiller, and five widely-used uncertainty quantification methods are compared under distribution shift scenarios. Both the model precision performance and robustness performance are investigated. The experimental result demonstrates that the deep ensemble (DE) model is a simple but efficient method to solve the distribution shift problem, where its precision performance and robustness performance outperformed the other methods, fol-lowed by Bayesian neural network (BNN). In addition, the benefit of robust deep learning model is also studied. It is shown that the deep learning model with higher robustness performance can better filter out the unreliable predictions, which is essential to the online active learning strategy. The proposed research framework might provide a solid foundation for the real-world application of deep learning model.

Automated summary

Artificial intelligence has become a key technology in building energy conservation, particularly in complex building energy systems modeling. However, existing deep learning models lack robustness under distribution shift scenarios, limiting their real-world application. To address this, a cluster-based dataset splitting method is proposed to simulate distribution shift scenarios, and uncertainty quantification methods are adopted to improve overall model robustness. Experimental results demonstrate that the deep ensemble (DE) model outperforms other methods in terms of precision and robustness, followed by Bayesian neural network (BNN). The research framework provides a solid foundation for the real-world application of deep learning models.