Safe and energy-efficient wide-range gas turbine control via condition-adaptive reinforcement learning

With the increasing demand for flexible operation in modern power systems, gas turbines are frequently required to operate under wide load ranges and varying environmental conditions. In practical operation, compressor surge margin is strongly influenced by operating conditions and environmental disturbances, which makes accurate surge-margin regulation challenging. Consequently, industrial control systems typically adopt conservative anti-surge strategies with large safety margins. These strategies ensure operational safety while causing substantial energy efficiency loss. A hypernetwork-based Twin Delayed Deep Deterministic Policy Gradient (TD3-Hyper) control framework is proposed for active surge-margin regulation under varying operating conditions. The proposed method generates condition-specific policy parameters in a dynamic manner, which enables adaptive control across different gas turbine operating regimes. A dual-boundary safety–efficiency reward formulation is introduced for simultaneous enforcement of surge safety and energy efficiency requirements. A five-stage curriculum learning strategy is incorporated for improved training stability and convergence behavior. The proposed approach enhances adaptability under wide operating conditions and improves adaptability and training stability for reinforcement learning-based gas turbine control under wide operating conditions. Compared with Adaptive MPC, it achieves a 1.27% reduction in cumulative fuel consumption under variable-load transient conditions. Robust performance is also observed under unseen operating conditions, together with consistent energy-saving behavior enabled by active surge-margin regulation, indicating its potential for safe and energy-efficient gas turbine operation.