GAN-Based Augmentation for Addressing Data Scarcity problem in Building Energy Load Prediction Using Optimized XGBoost

Predicting Heat Load (HL) and Cool Load (CL) of residential buildings at the early construction stage is a primary challenge of engineers and designers. Machine Learning (ML) tools bring the opportunity to predict the amount of building energy usage However, their performance is highly dependent on the amount of available data for training. In this study, only 768 samples were available. To overcome data scarcity problems and improve model prediction performance, a Tabular module based on Generative Adversarial Networks (TG) was proposed for synthetic training data generation, which generated 3863 new samples for training. The performance of the proposed TG based state-of-the-art eXtreme Gradient Boosting (XGBoost or XGB) Model tuned with Tree-Structured Parzen Estimator (TG-XGB-TPE), was evaluated on both original dataset and newly generated synthetic dataset. The results demonstrated that the proposed TG-XGB-TPE model outperformed the basic XGB and XGB-TPE. Compared to the basic XGB, it improved HL and CL estimation by 19% and 43%, respectively. Furthermore, it improved XGB-TPE model performance by 5.9% and 6.1% in HL and CL, respectively. Therefore, by using data augmentation techniques, the problem of data scarcity can be overcome, model exploration can be enhanced, and estimation performance can be improved. As a result, the developed model could be used by engineers and designers at the early stage of residential building construction to select the best plan among different plans from energy usage perspective with greater reliability.