面向电网边缘设备的端到端智能网络防御：方法及硬件在环仿真测试平台实现

This paper proposes and validates the operation of an end-to-end machine learning-based cyber defense framework for grid edge devices (GEDs) in distribution systems. The proposed framework establishes redundant communication paths between the GEDs and the utility control center at different levels of communication. Different control and measurement messages are duplicated and transmitted via these redundant paths. The collected duplicated messages are then aggregated using a novel machine learning-based data combination approach that is able to reduce or remove the effect of false data injection attacks (FDIA) on the communication network. The proposed framework defends against FDIA on the data transferring between the GEDs and the control center in any level of communication, including the edge level (i.e., GED communicates to a local controller) and the system level (i.e., local controllers communicate to the utility's central controller). An Attention-based Long-Short Term Memory (ALSTM) is proposed to build a reference model that enables combining the received redundant data. The proposed approach is implemented on the IEEE 33-bus test system running on a real-time Hardware-In-the-Loop testbed. The proposed approach is validated on an actual battery storage system, as part of the test system, while communicating to its local controller via multiple redundant paths. The results illustrate that the proposed framework significantly reduces the impact of FDIA on the limited number of redundant communication paths in real-time operation.