International Journal of Science, Engineering and TechnologyAuthors: Yogesh Ramesh Patni, Deepak P. Kadam, Kirti S. Kulkarni, Nilesh P. Dabe
Abstract: Real-time detection, classification, and localization of electrical faults are essential for fast protection and reliable operation of power systems. This paper presents a Diffusion-Enhanced Transformer (proposed hybrid model) that fuses a ResNet-like feature extractor and Transformer-based sequence learner with a diffusion-model generative module for data augmentation and robustness. The model is evaluated on simulated IEEE-9 bus fault waveforms and benchmarked against conventional CNN, LSTM and Transformer baselines.Experimental results demonstrate the framework’s strong real-time performance: per-fault detection accuracies of 99.1% (LG), 98.7% (LL), 98.3% (LLG) and 98.9% (LLL); overall classification metrics with precision/recall/F1 around 98.9% / 98.6% / 98.7% for the proposed model and confusion matrix showing diagonal values >0.97). The proposed hybrid achieves 98.6% overall accuracy in comparative tests while reducing fault-location error to 1.52 km (MAE), and 1.97 km (RMSE), outperforming ResNet and LSTM baselines. These findings verify that incorporating diffusion-based generative augmentation with a Transformer backbone yields improved generalization on sparse/high-noise fault data, faster inference than standard Transformers, and more accurate localization than conventional deep models, making the approach suitable for deployment in smart substations and real-time protection schemes.
DOI: http://doi.org/10.5281/zenodo.20787331
