中圖分類號：TM615；TP393 文獻標志碼：A DOI: 10.16157/j.issn.0258-7998.256419
中文引用格式： 周旺. 光伏功率預測的對抗攻擊與防御研究[J]. 電子技術應用，2025，51(12)：115-119.
英文引用格式： Zhou Wang. Research on adversarial attack and defense of photovoltaic power prediction[J]. Application of Electronic Technique，2025，51(12)：115-119.

Research on adversarial attack and defense of photovoltaic power prediction

Abstract： Deep neural networks have been widely used in photovoltaic power prediction, but they are vulnerable to adversarial attacks. In order to improve the robustness of the prediction model, an adversarial attack algorithm based on fast gradient sign method and a defense algorithm based on adversarial training are proposed. The fast gradient sign method generates adversarial samples with temporal correlation, and establishes the quantitative relationship between attack intensity and prediction error. Adversarial training enhances the generalization ability of the model to input disturbances by combining adversarial samples to resist adversarial attacks. Experimental data show that adversarial attacks can significantly reduce the prediction accuracy of the model, and the model trained by adversarial training can effectively improve the robustness. This method verifies the effectiveness of the countermeasure attack and defense mechanism in photovoltaic power prediction, and has practical application value for the safe operation of power system.

Key words : photovoltaic power prediction；adversarial attack；fast gradient sign method；adversarial training