CWIA: A Model-Free, Low-Cost Adversarial Perturbation Method for Wind Power Forecasting
Adversarial robustness assessments of wind power forecasting should account for attacker resources, cross-architecture stability, and detectability. Existing attacks commonly require surrogate-model preparation or perturbation learning, leaving the threat from minimal-preparation attacks insufficiently characterized. We propose the continuous wind injection attack (CWIA), a deterministic method that directly perturbs wind speed using an available regional input scale. CWIA requires no private wind farm training data, surrogate training, target predictions or gradients, universal-perturbation fitting, or online optimization. A predefined temporal ramp confines injection to wind speed and permits online application through a single vector addition with boundary projection, retaining the additive application of a prepared UAP. Evaluations on PSML use four forecasting architectures, six surrogate-based attack baselines, constant-offset ablations, and nine anomaly detectors to examine effectiveness, preparation requirements, and temporal boundary behavior. On MISO, CWIA produces MAEs of 0.1335–0.1443 p.u. across the four targets at a 15% normalized maximum budget, while its zero-start ramp preserves the tested temporal boundary score, establishing a useful low-preparation baseline with detector-dependent detectability.
Authors
- Lu Wei (ORCID: https://orcid.org/0000-0001-6120-4937)
- Zheng Qian (ORCID: https://orcid.org/0000-0002-2657-6361)
- Liliang Wang (ORCID: https://orcid.org/0009-0001-5897-230X)
Institutions
- Beihang University (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-13
- DOI
- https://doi.org/10.3390/app162211252
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00