Control Optimization of Steam Boilers via Reinforcement Learning

Steam boilers present complex control challenges due to their nonlinear dynamics and fluctuating load conditions. This study proposes a novel hybrid adaptive control framework that significantly enhances system performance by synergizing off-policy deep reinforcement learning, model reference adaptive control (MRAC), and optimized PID control through a weighted fusion strategy. Three innovative hybrid controllers, Weighted-Fusion-DDPG-MRAC-PID, Weighted-Fusion-TD3-MRAC-PID, and Weighted-Fusion-SAC-MRAC-PID, are developed and rigorously benchmarked against standalone MRAC and optimized PID for precise steam flow and drum pressure regulation in a two-state boiler system under four diverse operational conditions. The adaptive fusion mechanism dynamically modulates control actions (feedwater flow and applied heat) in response to system variations, ensuring superior stability. Furthermore, a novel nonlinear settling time algorithm is introduced to accurately assess transient responses in complex signals, addressing the limitations of conventional methods. Extensive experimental validation confirms the superiority of hybrid controllers, achieving significantly faster settling times, peak overshoot reductions to $(\leq 2\%)$ , and lower error metrics than MRAC alone. While optimized PID serves as a baseline, hybrid controllers consistently achieve faster settling times and improved robustness under most nonlinear operating conditions. Weighted Fusion-DDPG-MRAC-PID offers the best trade-off between performance and computational efficiency. These findings underscore the practical flexibility of the proposed framework for advanced steam boiler control.