A Physics-Informed Graph Reinforcement Learning Framework for Robust Coordinated Control of Sub-Synchronous Oscillations in Weak Grids with Multi-Machine Inverter-Based Resources

The rise of inverter-based resources and converter-interfaced renewable energy systems has changed the dynamic traits of modern power systems, increasing susceptibility to sub-synchronous oscillations and resonance under weak-grid conditions. These oscillations stem from intricate interactions between converter control loops, network impedance, electromechanical dynamics, and control mechanisms, which can lead to poorly damped modes, instability, and large-scale disturbances. Conventional mitigation methods, such as damping controllers, adaptive filtering, FACTS compensation, and optimization tuning, often face issues like model dependency, fixed parameters, limited operating ranges, and poor coordination among distributed IBRs in uncertain conditions. This paper proposes a Physics-Informed Graph Reinforcement Learning (PI-GRL) framework to adaptively and coordinately mitigate SSOs in multi-machine IBR-dominated weak power systems. The proposed framework is validated on a modified IEEE 39-bus system with synchronous generators, DFIG/PMSG wind farms, and hybrid grid-forming/grid-following converters. High-resolution PMU data, including voltage, current, and power measurements, support data-driven monitoring and control.

The PI-GRL framework combines a Physics-Informed Graph Neural Network (PI-GNN) with multi-agent reinforcement learning (MARL) and a H∞ robust control layer. The PI-GNN incorporates power-flow constraints, swing dynamics, converter control relationships, and network topology into a graph, allowing for precise identification of critical oscillatory modes and overall system stability characteristics. Distributed MARL agents coordinate damping control by adapting converter parameters, virtual impedance settings, and current injection references based on graph features. The framework uses decentralized agent coordination and lightweight inference for real-time deployment with minimal computational load. The H∞ robust control layer adds stability and resilience against parameter uncertainty, measurement noise, and external disturbances. The proposed approach is assessed across various operating scenarios, such as low short-circuit ratio conditions (SCR < 2), different series-compensation levels (30–70%), renewable penetration above 70%, random wind-speed and load fluctuations, three-phase faults, post-fault recovery, and mixed grid-forming/grid-following interactions. Performance assessment involves eigenvalue analysis, damping-ratio evaluation, participation-factor analysis, impedance-based stability assessment, and electromagnetic transient simulations.

Simulation results show that the PI-GRL framework greatly improves oscillatory stability, increasing critical-mode damping ratios from about 0.03–0.06 to 0.18–0.32, with average enhancements of 250–400%. Oscillation settling times drop by 55–70%, and peak oscillation amplitudes fall by 60–80% compared to traditional damping-control methods. Comparative studies show that GA/PSO-optimized controllers, battery-energy-storage-based damping methods, adaptive multi-modal damping controllers, and reinforcement-learning-based supplementary damping controllers achieve 30–50% faster oscillation suppression, 25–45% greater damping enhancement, and improved robustness in various operating conditions.

The framework effectively addresses coupled oscillatory mechanisms, such as induction-generator effects, torsional interactions, and converter-control-induced oscillations. The generalization capability is shown across various operating conditions and disturbances not directly faced during training, yet within the range of configurations and scenarios considered during training and validation. The proposed PI-GRL framework offers a scalable and effective solution for coordinated SSO mitigation and stability improvement in renewable-dominated weak-grid settings.