Physics-informed generative adversarial networks for reliable long-term harmonic forecasting in offshore wind farms

Abstract

This study proposes a Physics-Informed Generative Adversarial Network (PI-GAN) framework for reliable long-term harmonic distortion forecasting in offshore wind farms. Unlike conventional GANs, the proposed model embeds domain-specific constraints into generator training by enforcing current-voltage proportionality and wind speed-dependent harmonic behavior. The framework is evaluated in a long-term setting by training on data from 2005-2009 and testing on November 2019, thereby introducing a 10-year temporal gap. To examine the effect of physics-informed synthetic data generation across different predictors, both machine learning and deep learning backbones are considered, including Random Forest, XGBoost, GRU, TCN, DLinear, MLP-Mixer, and DRFormer models. Results show that PI-GAN preserves competitive statistical accuracy while reducing physics-violation errors by approximately 45%-50% relative to standard GAN-based baselines across all evaluated backbones. Notably, even state-of-theart architectures such as DLinear, MLP-Mixer, and DRFormer produce physically inconsistent forecasts when paired with a standard GAN, but achieve a 46%-47% reduction in physics violations when the same architectures are used with PI-GAN, demonstrating that physical plausibility is determined by the data generation process rather than by the forecasting model itself. These findings indicate that the principal benefit of PI-GAN is not merely lower conventional prediction error, but the generation of forecasts that are more physically plausible and interpretable. Therefore, PI-GAN provides a more reliable data-generation and forecasting framework for renewable energy systems.