Dual-Objective Adversarial Attack on Wind Power Forecasting: Balancing Destructiveness and Stealthiness

As inputs to wind power forecasting models, meteorological factors are vulnerable to adversarial attacks, leading to deviations in forecasted wind power. Existing research on adversarial attacks against renewable energy forecasting models mostly focuses on a single destructive objective, neglecting attack stealthiness. This makes it challenging to achieve actual destructive effects in scenarios where example detection mechanisms are deployed. Therefore, this paper proposes a Dual-Objective Adversarial Attack (DOAA) algorithm that balances attack destructiveness and stealthiness to achieve the joint optimisation of both. First, a wind power forecasting model based on deep neural network is constructed, which accurately captures the complex coupling relationships between meteorological factors and wind power. Second, a graph autoencoder (GAE) model suitable for time-series example detection is designed, which fuses the temporal correlations of data from adjacent time periods using graph structures and realises the detection of adversarial examples through reconstruction loss. Finally, the above models are integrated to construct the DOAA algorithm based on Projected Gradient Descent. By adjusting the coefficient of the reconstruction loss term, the preference for attack destructiveness or stealthiness can be achieved. Experimental results verify that the proposed DOAA algorithm achieves a favourable balance between attack destructiveness and stealthiness.