AugPlug: An Automated Data Augmentation Model to Enhance Online Building Load Forecasting

Yang Deng, Rui Liang, Yaohui Liu, Jiaqi Fan, Dan Wang

Abstract: Online Building Load Forecasting (BLF) is a scheme that designs a model update strategy to continuously update the deployed ML-based BLF model to adapt to changes in the distribution of data. Many online BLF schemes have recently been developed. However, updates can be ineffective, resulting in a decay in accuracy or even in performance that is worse to compared to that without the update. One primary reason for this is poor preparation of the data used to update the model (namely the updating set), since most of the online BLF schemes that have been developed update the ML model using collected historical data, which may not reflect the characteristics of the future distribution of data.To prepare a suitable updating set for the BLF model update is a challenging and ad hoc exercise. In this paper, we propose to leverage automated data augmentation (AutoDA), a data augmentation (DA) framework based on reinforcement learning, to automatically search for the optimal DA policy to generate synthetic data. We thus develop AugPlug, a data augmentation model to instantiate AutoDA in online BLF and demonstrate how it can generate updating sets. A unique advantage of AugPlug is its plug-and-play compatibility for integration into different online BLF schemes. Comprehensive experiments on four published online BLF schemes, involving hundreds of buildings, show that AugPlug can improve the overall performance of the online BLF by 29.37%.

External IDs:doi:10.1145/3671127.3698190