Secure Solar Forecasting: Deep Learning Approaches for Cyber Attacks Detection and Mitigation

• Published in: 2024 6th International Conference on Smart Power & Internet Energy Systems (SPIES) pp. 206 - 211
• Main Authors: Abughali, Ahmed, Alansari, Mohamad, Al-Sumaiti, Ameena S., El Moursi, Mohamed S.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.12.2024
• Subjects: Cyber Security; Cyberattack; Deep learning; Denial of Services (DoS); False Data Injection Attack (FDIA); Forecasting; Internet; Long short term memory; Noise; Noise measurement; Prevention and mitigation; Security; Solar energy
• DOI: 10.1109/SPIES63782.2024.10983606

Due to the increasing integration of solar energy into the energy infrastructure, solar power forecasting has become an attractive target for malicious attackers. However, existing studies in the literature primarily focus on detecting a single type of cyber attack, namely False Data Injection Attacks (FDIAs), and they lack recovery approaches. This paper proposes two deep learning-based security schemes: a two-stage detection approach and a mitigation approach. The proposed schemes are designed to accurately detect and recover stealthy FDIAs and Denial of Service (DoS) attacks. The first stage in the detection model determines whether the sample has been attacked, while the second stage categorizes the type of attack if an attack is present. After identifying the attack type, the proposed mitigation model recovers the corrupted measurements. For comparative analysis, the performance of three different deep learning models is evaluated for both tasks: detection and mitigation. The proposed models are tested utilizing a real Global Horizontal Irradiance (GHI) dataset collected from Abu Dhabi between 2017 and 2019. The two-stage detection approach yielded significantly better results than tackling the problem directly as a multi-class classification task. Specifically, the best-performing two-stage detection model, the Long Short Term Memory (LSTM) model, showed an average improvement of 6.52% across the Area Under the Curve (AUC). Additionally, the best-performing proposed mitigation method, LSTM, substantially recovered the corrupted measurements by 92.53% for FDIAs and 99.62% for DoS attacks.