Wide-area event classification in power systems using EfficientNet with robust adversarial defence

Resilience and security of the smart-grid are dependent on wide area monitoring, protection and control (WAMPC) systems, which are fueled by the data generated by high-resolution phasor measurement units (PMUs). Recently, Deep learning (DL) has shown strong potential for automating the detection and classification of disturbances based on PMU datasets. This work proposes EfficientNet-based DL frameworks for supervisory protection and event diagnosis in the transmission network. In contrast to conventional convolutional neural networks, EfficientNet employs a compound scaling approach that balances model depth, width and resolution to achieve high accuracy with fewer parameters. Here, the pseudo-colour image representation of spatio-temporal PMU data is used to train the EfficientNet models for classifying various power system disturbances, including faults, generation loss and synchronous motor switching events. In this regard, the three different versions of the EfficientNet model, such as EfficientNet-B0, B1 and B2, are tested. The experimental results demonstrate improved classification performance compared to traditional DL architectures. Moreover, the work evaluates the robustness of the proposed scheme under adversarial scenarios by crafting black-box attacks using DeepFool perturbations. EfficientNet-B2 achieved a high baseline accuracy (0.9901 ± 0.0121) in five-fold cross-validation, demonstrating its potential to enhance the reliability and scalability of DL-based WAMPC systems.