A novel deep generative modeling-based data augmentation strategy for improving short-term building energy predictions

Short-term building energy predictions serve as one of the fundamental tasks in building operation management. While large numbers of studies have explored the value of various supervised machine learning techniques in energy predictions, few studies have addressed the potential data shortage problem in developing data-driven models. One promising solution is data augmentation, which aims to enrich existing building data resources for reliable predictive modeling. This study proposes a deep generative modeling-based data augmentation strategy for improving short-term building energy predictions. Two types of conditional variational autoencoders have been designed for synthetic energy data generation using fully connected and one-dimensional convolutional layers respectively. Data experiments have been designed to evaluate the value of data augmentation using actual measurements from 52 buildings. The results indicate that conditional variational autoencoders are capable of generating high-quality synthetic data samples, which in turns helps to enhance the accuracy in short-term building energy predictions. The average performance enhancement ratios in terms of CV-RMSE range between 12% and 18%. Practical guidelines have been obtained to ensure the validity and quality of synthetic building energy data. The research outcomes are valuable for enhancing the robustness and reliability of data-driven models for smart building operation management.