Conditional generative adversarial network (cGAN) for generating building load profiles with photovoltaics and electric vehicles

Abstract

Building load profiles are essential in research on building energy management, efficiency, demand response, and grid planning. With the growing adoption of solar photovoltaics (PV) and electric vehicles (EVs), integrated building load profiles are becoming increasingly important for effective management. However, current methods for generating building load profiles focus only on buildings without considering PV and/or EV adoptions. To address this gap, we propose using the conditional generative adversarial network (cGAN), a machine learning technique that generates realistic data conditioned on specific inputs, to create building load profiles that account for PV and EVs. This approach was tested using a smart meter dataset from a major metropolitan area in the southwest United States, containing years of hourly readings from 110 households with PV and EV adoptions. We extracted the key parameters that can describe the generated and real load profiles, and compared their mean and standard deviation to validate the results. KL divergence and FID scores were also used to compare the distributions. The results showed strong alignment between the generated and actual smart meter data across all PV, EV and seasonal conditions. The data under different combinations of PV, EV and weather conditions serve as conditional inputs for the cGAN, allowing it to generate building load profiles that maintain key statistical characteristics for both cooling and heating seasons, and various installation status of PV and EV. Additionally, this method safeguards customer privacy and reduces the effort needed for analyzing occupant behavior and building physics, which are typically required in physics-based energy models.