A Gradient-Based Wind Power Forecasting Attack Method Considering Point and Direction Selection
Runhai Jiao, Zhuoting Han, Xuan Liu, Changyu Zhou, Min Du
Abstract
Machine learning methods have been prevailing in wind power forecasting, while these data-driven based methods are susceptible to cyberattacks. Typical attack methods inject malicious data into influence factors according to the gradient direction of the forecasting model to randomly increase or decrease forecasting results, ignoring the number of attacks and attack effect. In this paper, an attack sample selection model is proposed to select vulnerability sample points for attack in order to reduce the number of attacks. At the same time, an attack direction judgment model is developed to launch the attack in the correct gradient direction to maximize the attack effect. Moreover, the effectiveness of the proposed approach is validated on two public wind power datasets and nine typical machine learning based forecasting models such as ANN, ENN, RNN, LSTM, GRU, BiLSTM, BiGRU, CNN and TCN. Compared with the existing gradient-based attack methods, the proposed attack method increases MAPE values of the nine models by about 9% on average while improving the attack concealment.