面向直流微电网的PINN混合网络防御机制：抵御重放/虚假数据注入攻击的弹性可靠控制

With the growing reliance on DC microgrids (DC MGs) in critical infrastructure, securing them against sophisticated cyberattacks is essential. This study presents a hybrid cyber-defense (HCD) framework that detects and mitigates false data injection (FDI) and replay attacks through a combination of bidirectional LSTM (Bi-LSTM) autoencoders for anomaly detection, cross-correlation analysis for replay attack identification, and a physics-informed neural network (PINN) for adaptive control. A Kalman filter-based estimator maintains control stability when sensor measurements are compromised. The framework operates in a fully decentralized manner at the local controller level, enabling fast, low-latency responses. Validation on a 4-bus MATLAB/Simulink model and a 3-bus DC MG experimental testbed demonstrates over 95% detection accuracy and effective mitigation, preserving voltage stability under attack. The proposed control algorithm is deployed in the AT19SAM3X8E microcontroller, acting as the local controller at each node, occupying 40 $KB$ of memory with an execution time of 0.4 $ms$. This real-time deployment confirms practical applicability for lightweight, standalone operation. These results demonstrate the proposed method’s robustness and scalability, advancing intelligent, real-time cyber-resilience for future DC MG.