<p>The introduction of Internet of Things (IoT) devices into smart grid frameworks has posed significant cybersecurity challenges, including advanced attacks such as false data injection, distributed denial-of-service (DDoS), and sophisticated network intrusions. In this paper, a new hybrid deep learning architecture is proposed that combines Convolutional Neural Networks (CNNs) to extract spatial features with Long Short-Term Memory (LSTM) networks to capture temporal patterns, and an autoencoder-based preprocessing unit to reduce dimensionality and remove noise. The architecture suggested, AE-CNN-LSTM, can detect anomalies in real time and is designed as an edge-deployable framework, making it highly suitable for resource-constrained IoT-enabled smart grid environments. The experimental validation of the technology on two benchmark datasets, CICIoT2023 and UNSW-NB15, shows improved performance, with accuracies of 99.15% and 98.87%, respectively, surpassing the state-of-the-art. To establish the statistical significance of our results (<i>p</i> &lt; 0.05), we performed paired t-tests. Extensive ablation experiments confirm the usefulness of individual architectural building elements, and experiments with multiclass classification show that it is possible to identify seven distinct attack types. The architectural design of the autoencoder component provides inherent robustness to distribution shifts, offering potential resilience against previously unseen attack patterns, though explicit zero-day evaluation remains a direction for future work. The framework has an inference latency of 2.8ms per sample and can be deployed on an edge computing platform with limited resources to support next-generation smart grid security.</p>