Volume 42: Energy Transitions toward Carbon Neutrality: Part V

DAF-GAN: Day-Ahead Forecasting of Building HVAC Energy Consumption Using Multi-Channel Generative Adversarial Networks Yichuan X. Ma, Lawrence K. Yeung

Abstract

In the pursuit of sustainability and energy efficiency, accurate short-term prediction of HVAC energy consumption is crucial. Deep learning emerges as a promising solution for handling diverse data challenges in building HVAC systems. While deep generative learning excels in computer vision, its potential in predicting energy consumption remains largely untapped. This study first introduces a novel framework, transforming forecasting into a conditional generative problem in the temporal domain. We then propose DAF-GAN, an image inpainting-based data-driven method for Day-Ahead Forecasting of buildings’ HVAC energy consumption using multi-channel Generative Adversarial Networks (GANs). In day-ahead forecasting tasks across eleven real-world buildings, DAF-GAN exhibits relative improvements of 17% to 68% across four different error metrics compared to six traditional and deep learning models. DAF-GAN also demonstrates less bias and superior stability when applied to different buildings, holding promise for enhancing energy-efficient building automation and management.

Keywords building energy forecasting, HVAC, conditional generative adversarial network, deep learning, image inpainting, artificial intelligence