Adaptive TD3-PID control for once-through steam generator using hardware-in-the-loop simulation

This paper presents a adaptive hybrid control framework that integrates Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning with PID. This approach addresses the critical challenges of pronounced non-linearity, significant time delays, and wide-ranging operating conditions in the pressure control of Once-Through Steam Generators (OTSGs). By employing a TD3 agent to dynamically optimize PID within predefined safety bounds, the proposed strategy substantially improves system adaptability without compromising the inherent stability of the baseline PID structure. Extensive hardware-in-the-loop (HIL) experiments on an industrial distributed control system (DCS) demonstrate the effectiveness of the proposed strategy. Quantitative results show that the TD3-PID control reduces settling times by more than 50% during setpoint tracking. Moreover, under large-scale load transients, the intelligent agent suppresses pressure overshoot by 73%-90% and effectively attenuates associated temperature fluctuations. These results validate the robustness and practical feasibility of the proposed TD3-PID architecture, offering a viable pathway for the deployment of intelligent control systems in advanced nuclear energy applications.