Inverse Dynamic-Augmented Deep Reinforcement Learning for Set-point Control of Power Generation Systems
摘要
Set-point control of power generation systems remains a critical challenge due to their highly nonlinear dynamics and stringent operational constraints. Conventional control frameworks exhibit limited adaptability in handling transient disturbances and require extensive manual tuning of control parameters, which significantly restricts their applicability under complex operational scenarios. To bridge this technological gap, this paper introduces a novel inverse dynamic-augmented deep reinforcement learning (DRL) framework for set-point control. Specifically, long short-term memory networks are embedded within the DRL agent to capture long-range dependencies and cyclic operational patterns in power generation systems. Meanwhile, a multi-objective reward-shaping mechanism combining sparse and dense rewards is designed to coordinate short-term behavioral optimization and long-term goal achievement in policy iteration. In addition, a deep neural network-based inverse dynamic model is developed for feed-forward compensation, which mitigates model-plant mismatch through real-time compensation of nonlinear hysteresis effects. Comprehensive experimental validation on a power generation system demonstrates superior performance, achieving enhancements of 38.56