The increasing penetration of renewable energy
and inverter-based distributed generation has introduced
significant challenges for fault detection in modern smart grids.
Traditional protection schemes often struggle to identify weak
or evolving fault signatures, leading to delayed fault isolation
and compromising system reliability. To address this gap, this
study introduces a hybrid deep learning framework that
integrates convolutional neural networks (CNNs) with long
short-term memory (LSTM) networks for intelligent fault
detection. The approach leverages phasor measurement unit
(PMU) image data, where CNN layers extract spatial fault
characteristics and LSTM layers capture temporal dynamics,
enabling a more comprehensive representation of fault
progression. Data preprocessing included normalization, class
rebalancing, and synthetic noise augmentation to ensure
robustness. Model performance was validated using stratified 5-
fold cross-validation, achieving 99% classification accuracy
while maintaining lower computational requirements compared
to CNN-only and ensemble-based baselines. Comprehensive
evaluation with ROC–AUC, PR–AUC, per-class accuracy, and
confusion matrices further demonstrated reliability and
interpretability. The findings highlight the potential of the
proposed method to enhance fault detection mechanisms,
contributing to improved grid stability, faster protection
response, and greater resilience of future smart energy systems

The increasing connectivity of Electric Vehicle Supply Equipment (EVSE) within smart grid networks has compounded the threat of cyber exposures, particularly Distributed Denial-of-Service (DoS) attacks. This paper introduces a lightweight, real-time intrusion detection system based on a hybrid deep learning structure that intermixes Transformer encoders and a Multilayer Perceptron (MLP) classifier. The proposed model uses kernel-level event logs to capture both temporal dependencies and high-dimensional feature interactions. A well-structured preprocessing pipeline includes feature leakage prevention, normalization, class balancing, and stratified cross-validation. This approach ensures data integrity and effective learning. Experimental evaluation on a real-world dataset confirms the model’s superior performance with 100% accuracy, precision, recall, and F1-score, and ideal ROC-AUC and PR-AUC scores. Furthermore, the framework is streamlined for deployment on resource-constrained edge devices to enable decentralized, on-device threat detection. These results accentuate the effectiveness and viability of the suggested solution in enhancing the cybersecurity posture of modern EV charging ecosystems