ABSTRACT With the advancement of science and technology in power systems, protection devices for distribution equipment play a key role in safeguarding operations. However, they face security challenges such as misoperation induced by adversarial samples. Due to a lack of adaptability to time‐varying attack modes, traditional rule‐based or static defense mechanisms exhibit limited protective effectiveness. To address this issue, this study proposes a collaborative defense framework that integrates Digital Twin and Reinforcement Learning (RL) technologies. The framework constructs a virtual model of a distribution protection device, simulates multi‐dimensional attack scenarios in a virtual environment, and provides a training platform for the power analysis system. The designed dynamic adversarial training environment incorporates a physically constrained reward function, which ensures that defense decisions comply with power system safety boundaries while overcoming the feasibility challenges of deploying the algorithm on edge devices. Simulation results demonstrate that the misoperation rate is reduced to below 3.5% under sample attacks, and the attenuation index of generalization performance against unknown attack modes is maintained at a low level of 0.31. After lightweight processing, the inference delay of the model on embedded devices is controlled within 10 ms, meeting the real‐time requirements of power protection and offering a technical pathway for developing intelligent security protection solutions in power systems.
Zhang et al. (Mon,) studied this question.