Secure FDRL for safe decision-making and policy embedding validation in hybrid microgrids under hybrid cyberattacks

Hybrid AC/DC microgrids are more vulnerable to sophisticated cyberattacks with their growing reliance on digital control and communication infrastructures. This paper presents a cyber-resilient three-step framework to include minimum mean squared error filtering, federated deep reinforcement learning (FDRL) combined with soft actor-critic, and locality-sensitive hashing-based policy verification to detect, assess, and counter hybrid cyberattacks. The first step of the proposed framework deals with raw sensor data to find anomalies using the Mahalanobis distance to deliver a state of trusted representation. The second step involves using an FDRL policy learning paradigm, which is decentralized, such that data robustness and privacy are maintained. Policies learned at this stage are inserted as a vector of low-dimensional form and are compared using a trusted database using cosine similarity in the third step; policies scoring lower than the threshold similarity are blocked by safe fallbacks. The proposed method positively pre-filters adversarial control strategies before deployment to enhance operational resilience. Simulation on IEEE 69-bus and 141-bus (Caracas) test systems confirms increased detection accuracy (AUC ≈ 0.98), faster voltage recovery (<2.6 s), and a 30–55% improvement in control performance over baselines. This work introduces an extendable and secure architecture to safeguard microgrids against cyber-physical attacks.