Deep learning-based feature engineering methods for improved building energy prediction

The enrichment in building operation data has enabled the development of advanced data-driven methods for building energy predictions. Existing studies mainly focused on the utilization of supervised learning techniques for model development, while overlooking the significance of feature engineering. Feature engineering are helpful for reducing data dimensionality, decreasing prediction model complexity, and tackling the problem of corrupted and noisy information. Considering that each building has unique operating characteristics, it is neither practical nor efficient to manually identify features for model developments. Data-driven feature engineering methods are thus needed to ensure the flexibility and generalization of building energy prediction models. Using operation data of real buildings, this paper investigates the performance of different deep learning techniques in automatically deriving high-quality features for building energy predictions. Three types of deep learning-based features are developed using fully-connected autoencoders, convolutional autoencoders and generative adversarial networks respectively. Their potentials in building energy predictions have been exploited and compared with conventional feature engineering methods. The study validates the usefulness of deep learning in enhancing building energy prediction performance. The research results help to automate and improve the predictive modeling process while bridging the knowledge gaps between deep learning and building professionals.