A robust IoT architecture for smart inverters in microgrids using hybrid deep learning and signal processing against adversarial attacks

Abstract

The increasing autonomy and deployment of cyber-physical systems, particularly power electronics-based inverters within microgrids, has heightened their vulnerability to cyber threats, such as False Data Injection (FDI) and adversarial attacks, which can compromise the integrity of data exchanged across communication networks. To address these security concerns, this paper proposes a new Internet of Things (IoT) architecture that integrates a hybrid approach combining 2-D Convolutional Neural Networks (2-D CNN) with Continuous Wavelet Transform (CWT) for enhanced cyberattack detection. The framework is designed to detect and mitigate adversarial perturbations, focusing on FDI and other attack vectors targeting the communication infrastructure of smart inverters. By transforming raw data into images using CWT, the framework enables efficient statistical feature extraction, enhancing learning accuracy to approximately 98.9 %, outperforming other models. Additionally, it reduces the computational load of signal processing, achieving a processing time of just 0.0548 s. The proposed deep learning model is tested against various levels of cyber perturbations, and its performance is benchmarked against other deep learning and machine learning techniques. The framework is validated using real-time data from a practical distribution system equipped with smart inverters, demonstrating its effectiveness in safeguarding microgrids from cyber threats.