Concept Drift-aware Time-series Generation for Online Building Load Forecasting: An Automated Data Augmentation Paradigm

Online Building Load Forecasting (BLF) is a scheme that designs a model update strategy to continuously update the deployed machine learning (ML)-based BLF model, ensuring its adaptation to changes in data distribution. Numerous online BLF schemes have been proposed recently. However, updates may prove ineffective, leading to a degradation in accuracy or even performance inferior to that of models without updates. A primary cause of this issue is the inadequate preparation of the data used for model updates (referred to as the updating set ). Most existing online BLF schemes rely on historical data for model updates, which may fail to capture the characteristics of future data distributions. Preparing a suitable updating set is a challenging and often ad hoc task. In this paper, we propose the use of automated data augmentation (AutoDA) , a reinforcement learning-based data augmentation (DA) framework, to automatically search for optimal DA policies for generating synthetic data. We introduce AugPlug and demonstrate its capability to generate effective updating sets. A distinctive advantage of its plug-and-play compatibility enables seamless integration into various BLF schemes. Extensive experiments conducted on four published online BLF schemes, involving hundreds of buildings, demonstrate that AugPlug improves the performance of online BLF by 29.37%.