The application of data augmentation technology based on adversarial sample generation in distribution network acceptance testing

Distribution network acceptance ensures safety and reliability by verifying compliance with technical standards. Traditional methods are labor intensive, and data scarcity limits deep learning approaches. This paper proposes a Physics-Constrained Generator Network for adversarial sample generation, embedding power system laws to ensure physically valid and diverse samples. A dual discriminator framework evaluates authenticity and physical plausibility through multi-objective adversarial training. Experimental results demonstrate significant improvements in sample quality, diversity, and physical validity compared to existing methods. In practical applications, models trained with our generated samples achieved a 5.8% increase in acceptance accuracy, highlighting the method’s effectiveness.