• Autonomous digital twin framework for gas turbine combined cycle power plants. • Simulation-based comparison of conventional and reinforcement learning controllers. • Conventional control is safe but slow; reinforcement learning violates constraints. • Agent-augmented learning balances performance, safety, and resilience. • Autonomy assessed by adaptability, resilience, diagnostic accuracy, and robustness. Gas turbine combined cycle plants are expected to remain central to low carbon power generation while meeting electricity demand from artificial intelligence and data centers. To address these challenges, gas turbine combined cycle operation must evolve beyond rule-based control toward autonomy. This study develops a proof-of-concept autonomous digital twin for supervisory control loops in a gas turbine combined cycle plant that integrates reinforcement learning with supervisory agents and uses this digital twin to perform a simulation-based comparative evaluation of controllers under three representative scenarios: load ramping, fuel switching, and sensor faults. The digital twin is anchored to a publicly documented natural gas combined cycle reference plant from the United States National Energy Technology Laboratory, providing transparent steady-state validation. Forward validation is additionally performed using operational data from a Korean commercial Gas Turbine Combined Cycle plant. Proportional integral and reinforcement learning controllers are implemented as baselines, while reinforcement learning with agents incorporates modules for constraint enforcement, sensor arbitration, and transition stabilization. Results show that proportional integral control maintains safety but is slow and lacks fault tolerance, whereas reinforcement learning improves adaptability but violates turbine inlet temperature constraints and fails under sensor anomalies. Reinforcement learning with agents uniquely balances speed, safety, and resilience, achieving smooth load ramps without violations, stable fuel transitions, and reliable recovery under faulty sensing. Synthesizing these results into an autonomy framework defined by adaptability, resilience, diagnostic accuracy, and robustness, the study demonstrates that reinforcement learning with agents overcomes the performance–safety trade-off and provides a reproducible pathway toward autonomous gas turbine combined cycle plants.
Hwang et al. (Sun,) studied this question.