ENHANCEMENT OF TRANSFORMER DIFFERENTIAL PROTECTION USING PHYSICS-GUIDED SYNTHETIC DATA GENERATION AND HYBRID NEURO-SYMBOLIC CLASSIFICATION
Nitin Khandare, Prashant Gaidhane
Abstract: The ability to discriminate reliably between magnetizing inrush currents and internal fault currents is also a major issue in the protection of power transformers. The traditional harmonic restraint and signal-processing-based designs have been plagued by maloperation in modern low-loss core materials, current transformer (CT) saturation, noise, residual flux variation and current transformer (CT) saturation. Recent approaches based on artificial intelligence (AI) enhance accuracy, however, they are often restricted on massive datasets that are labeled well, which is problematic in real-world power systems and which do not take into account underlying physical limits. This paper presents a novel hybrid protection framework using a combination of physics-guided synthetic data generation, hybrid feature extraction, and neuro-symbolic classification to increase the benefit of transformer differential protection. To generate realistic inrush and internal fault current waveforms using a physics-guided generative adversarial network (PG-GAN), the models are physically constrained by restricting the electromagnetic behavior of magnetization and CT of the transformer. A hybrid collection of features, including physics-inspired indices, self-supervised deep temporal features and graph-based phase coupling descriptors is obtained. A neuro-fuzzy classifier enhanced With evidential reasoning is used to make final decisions that offer classification and confidence scores. The obtained simulation outcomes reveal higher security, dependability, and resistance to CT saturation, noises, and variability in the operating conditions than the traditional harmonic restraint and deep learning only approaches.
Keywords: Evidential reasoning; Graph Neural networks; Inrush current; Internal fault; Neuro-fuzzy classifier; Physics-guided GAN; Power transformer protection
DOI: 10.24874/PES08.03.034
Recieved: 21.05.2026 Revised: 06.07.2026 Accepted: 06.08.2026
UDC:
Reads: 1 
