Conditional Wasserstein GAN for Energy Load Forecasting in Large Buildings

Energy forecasting is necessary for planning electricity consumption, and large buildings play a huge role when making these predictions. Because of its importance, numerous methods to predict the buildings' energy load have appeared during the last decades, remaining an open area of research. In recent years, traditional machine learning techniques such as Random Forest, K-Nearest Neighbors, and AutoRegressive Integrated Moving Average (ARIMA) have been replaced with deep learning methods, which have an increased ability to capture underlying consumption trends. However, large amounts of data are mandatory for training neural networks to forecast energy load. With scarce data, augmentation techniques are necessary to ensure high-quality predictions. This paper introduces cWGAN-GP-SN, a conditional (convolutional) Wasserstein Generative Adversarial Network with Gradient Penalty and Spectral Normalization used to generate new electrical records. Our architecture leverages the advantages of multiple GAN models to enrich training stability and data quality. The experimental results based on the Building Data Genome dataset show how classification and regression tasks benefit from the enrichment of the dataset. Additionally, adversarial attacks were performed to investigate whether models trained on large amounts of synthetic data are more robust.