Intelligent Fault Diagnosis For Automated Power System Protection

Abstract

Modern power grids are getting more and more complex, which influences the need of intelligent, fast-response 
protection systems providing detection and classification of faults with low latency. In this paper, we deal with an 
empirical approach to design an automated power system protection scheme utilizing (ML) algorithms with an 
emphasis on comparative data across different classifiers architectures. This paper presents a multimodal dataset 
containing 32,550 samples collected from IEEE 39-Bus simulations, PSCAD-generated waveforms, MATPOWER 
benchmark cases, and actual Powergrid Corporation of India Limited (PGCIL) utility recordings, covering eight 
fault types and different voltage levels (11 kV to 400 kV). The features used in this paper were extracted in the 
time and time-frequency domain (with DWT and Fortescue symmetrical components) and selected using 
Recursive Feature Elimination (RFE). We train five ML models Random Forest (RF), Support Vector Machine 
(SVM), XGBoost, CNN-LSTM, and Transformer-based architecture so they are rigorously assessed under the 
same cross-validation condition. The CNN-LSTM hybrid model established in the research achieved 98.2% fault 
detection accuracy, 98.0% precision, and 12.6 ms mean detection latency, outperforming all traditional and state
of-the-art baselines. Data normality was established and validated for statistical significance using one-way 
ANOVA (F = 48.3, p < 0.001) and Tukey's HSD post-hoc test. These results demonstrate that a deep learning
aided signal-processing-feature combination is a significant step forward in automated protection scheme 
performance relevant to smart grid, including real-time relay coordination.