<p>This article presents an innovative method based on hybrid deep learning for real-time fault detection in smart grids, addressing the critical need for accurate and efficient fault identification in modern power systems. The proposed architecture integrates Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and attention mechanisms to capture spatial and temporal patterns in sensor data effectively. Comprehensive evaluations of the model demonstrate exceptional performance across multiple metrics, achieving an accuracy of 99.58% with the Hybrid CNN-LSTM-Attention model, significantly outperforming traditional approaches. The model exhibits high precision, recall, and F1 scores of 0.99, alongside a ROC AUC of 1.00, indicating its robust ability to distinguish between actual faults and non-fault conditions while minimizing false positives. This capability is essential for maintaining the stability of smart grid operations and preventing unnecessary outages. The results indicate that the hybrid approach enhances fault detection accuracy and provides a scalable and reliable solution for real-time monitoring in complex environments. This research contributes to the ongoing development of intelligent energy systems, offering valuable insights into applying advanced machine learning techniques for improved operational efficiency and resilience in smart grids.</p>