Publication

Abstract : Contemporary power grids exhibit unprecedented complexity due to renewable energy proliferation and dynamic load fluctuations, creating substantial challenges for conventional fault detection systems characterized by sluggish response mechanisms and constrained accuracy parameters. This investigation addresses the imperative requirement for enhanced fault identification capabilities through the development of an optimized deep learning framework utilizing EfficientNetB3 architecture integrated with Phasor Measurement Unit (PMU) data streams. The enhanced framework incorporates compound scaling methodology, sophisticated regularization techniques, and adaptive learning rate scheduling mechanisms to achieve superior classification performance across four distinct fault categories: normal operational states, generation loss events, synchronous motor switching transients, and generic fault scenarios. Experimental validation demonstrates substantial performance enhancements with 96.8% overall accuracy, representing a 9% improvement over conventional EfficientNetB3 implementations utilizing a comprehensive dataset of 5,000 PMU samples. The proposed methodology achieves exceptional F1-scores ranging from 94.4% to 97.8% across all fault categories, while maintaining computational efficiency with reduced training duration of 4.2 hours and inference latency of 6 milliseconds. These outcomes demonstrate the framework's potential for real-time deployment in smart grid monitoring infrastructures, offering considerable improvements over traditional SCADA-based approaches and contemporary machine learning methodologies.