Transmission Disturbances Cause Classification using COMTRADE Data Feature Extraction and Deep Neural Network Based on “Attention”

Abstract

This research presents a power system disturbance classification approach using a Deep Neural Network (DNN) architecture with an “attention” mechanism to analyze waveform data in COMTRADE format. The developed system classifies four types of transmission line disturbances, STRANGE OBJECTS (other objects), ANIMALS (animal disturbances), LIGHTNING and VEGETATION (trees/vegetation). The dataset consists of 849 successfully processed samples from 947 COMTRADE files, resulting in a processing success rate of 86.80%. The research methodology involves comprehensive feature extraction from time-series signals, generating 19,226 initial features, which were subsequently reduced to 150 optimal features through a multi-method feature selection process. The model architecture employs a transformer-inspired deep neural network with residual connections, layer normalization and tiered dropout to enhance generalization. The training process utilizes a combination of Focal Loss and Cross-Entropy to address class imbalance, alongside data augmentation and Test-time Augmentation (TTA) techniques to improve robustness. The model achieved its best performance with a balanced accuracy of 92.44% on validation data and 92.44% on test data, a Cohen’s Kappa of 0.8946, and a Matthews Correlation Coefficient (MCC) of 0.8963. The results indicate that the non-TTA approach with a separator provides the most optimal performance for deployment purposes. Class-specific performance was as follows: VEGETATION achieved the highest F1-score (0.974), followed by ANIMALS (0.914), STRANGE OBJECTS (0.911), and LIGHTNING (0.894). These findings suggest that the proposed model has the potential to support faster, more accurate, and more consistent disturbance cause diagnosis, thereby contributing to the enhanced reliability of the transmission system.