Momentum-Based Adversarial Attacks and Multi-Level Denoising Defenses in Deep Learning-Based Wind Power Forecasting.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Cybersecurity for Critical Infrastructure

Background:

• Deep learning models enhance wind power forecasting accuracy.
• These models are susceptible to adversarial attacks, causing significant forecast errors.
• Existing research often overlooks attack stealthiness, effectiveness, and defenses against multiple perturbation levels or in black-box settings.

Purpose of the Study:

• To propose a novel adversarial attack algorithm (MI-FGSM) for wind power forecasting.
• To develop a robust defense mechanism (MLI-DAE) against multi-level adversarial attacks.
• To evaluate the stealthiness and effectiveness of the proposed attack and defense strategies.

Main Methods:

• Developed the Momentum Iterative Fast Gradient Sign Method (MI-FGSM) incorporating momentum for generating adversarial samples.
• Proposed the Multi-Level Iterative Denoising Autoencoder (MLI-DAE) trained on adversarial samples with varying perturbation levels.
• Conducted experiments in both white-box and black-box scenarios to assess attack impact and defense efficacy.

Main Results:

• MI-FGSM demonstrated significantly higher forecast errors than FGSM, even with smaller perturbation magnitudes.
• MLI-DAE effectively restored attacked inputs to their clean forms across multiple perturbation levels.
• The proposed defense model maintained original forecast accuracy while mitigating adversarial attacks.

Conclusions:

• MI-FGSM presents a potent threat to deep learning-based wind power forecasting systems.
• MLI-DAE offers a robust and effective defense against sophisticated adversarial attacks.
• The study highlights the importance of addressing adversarial robustness in renewable energy forecasting.