Cyber-resilient control based on soft reinforcement learning for inverter-based PV/FC systems: Real-time implementation and validation

More recently, communication networks have been utilized in modern power systems for transferring measurement signals. While the utilization of communication networks adds a new layer to the system, the security of such systems is threatened by cyber-attacks. In such systems, appropriate detection is required to ensure the stability of grid-connected systems under such threats. In this paper, a defense scheme is developed to stabilize the voltage outcomes of an inverter-based integrated photovoltaic fuel cell (PV/FC) under cyber threats. The PV/FC case-study is configured by 150 kW/700 V PV radiation, a 150 kW/1400 V fuel cell, and a 265 kW three-level inverter. While the power system is threatened by cyber-attacks, a switching surface observer (SSO) is utilized in the control structure to estimate all perturbations and disturbances. An integral-backstepping controller (I-BSC) is proposed to compensate for any anomaly and stabilize outcomes of an integrated PV-fuel cell system. In order to capture cyber-attacks, a Soft Reinforcement Learning Agent (SRLA) with actor-critic neural network is adopted to adaptively adjust coefficients embedded in the established backstepping controller. In this framework, the disturbances and anomalies are estimated by SSO observer while the stabilization of voltage waveforms are reached by the adaptive I-BSC controller. Unlike conventional approaches, the proposed resilient scheme synergistically integrates real-time anomaly identification (via SSO), nonlinear stability-guaranteed tracking control (via I-BSC), and entropy-regularized adaptive policy learning (via SRLA) to ensure the system operation under disturbances and cyber-threats. The coordinated defense framework was evaluated via real-time simulation on an Arduino Mega2560 (ATmega-2560) platform. Verification of anomaly detection and compensation inclusive of robust tracking was tested across a variety of attacks including packet loss, ramp, and sinusoidal forms. Real-time outcomes of this work indicated significant improvements in stability, anomaly tolerance, and response accuracy under various scenarios.