Generative Adversarial Networks with Physics-informed Losses for High-resolution Load Profile Generation and Inpainting

Accurate load profile data are essential for optimizing energy systems. However, real-world datasets often suffer from low resolution and significant missing values. To address these challenges, this paper introduces physics-informed loss generative adversarial network (PIL-GAN), a model that combines generative adversarial networks (GANs) with physics-informed losses (PILs) derived from physics-informed neural networks (PINNs) that are integrated directly into the generator. High-resolution load profiles are generated that not only fill in missing data but also ensure that the generated profiles adhere to physical laws governing the energy systems, such as energy conservation and load fluctuations. By embedding domain-specific physics into the generation process, the proposed model significantly enhances data quality and resolution for low-quality datasets. The experimental results demonstrate notable gains in data accuracy, resolution, and consistency, making PIL-GAN an effective tool for energy management systems. The PIL-GAN also has broader applicability in other fields such as generating and inpainting high-resolution datasets for energy systems, industrial processes, and any domain in which data must comply with real-world physical laws or operational requirements.