Raw data.

Deep neural networks (DNNs) have achieved remarkable success in wind power forecasting, but DNNs are vulnerable to adversarial attacks that can severely degrade forecast accuracy. Existing studies primarily emphasize attack effectiveness and pay limited attention to attack stealthiness. In this paper, a dimension-constrained momentum iterative fast gradient sign method (DC-MI-FGSM) is proposed for wind power forecasting, which generates highly stealthy perturbations by applying the momentum update mechanism during attack optimization and limiting the perturbation dimensions of input samples. To defend against this attack, a denoising autoencoder (DAE)-based preprocessing defense strategy is developed for wind power forecasting, which resists adversarial attacks by mapping adversarial samples back to their corresponding clean forms. The effectiveness of the proposed attack and defense methods is validated on the public SDWPF dataset under both white-box and black-box scenarios. Compared with existing attacks, DC-MI-FGSM achieves a lower average perturbation percentage (APP), indicating superior attack stealthiness. Meanwhile, it causes more severe degradation in forecasting accuracy, as measured by MAPE, RMSE, and MAE, demonstrating stronger attack effectiveness. For defense, the proposed DAE-based preprocessing strategy effectively mitigates adversarial perturbations, significantly reducing forecasting errors while preserving the original accuracy on clean data. Moreover, it consistently outperforms adversarial training in terms of robustness and usability.