Reinforcement-Learning-Enhanced Nonlinear Model Predictive Control via Proximal Policy Optimization for the BioPower-5 Industrial Biomass CHP Plant

Residual oxygen content in the flue gas plays a key role in the efficient and clean operation of biomass-fired combined heat-and-power (CHP) plants. Because the flue-gas O2 concentration is governed by both the total air supply and the elemental composition of the fuel, it provides a reliable knowledge for the instantaneous combustion power. Consequently, state-of-the-art boiler control systems measure flue-gas O2 online and use it as a primary feedback variable. This paper proposes a novel non-linear model-predictive control (NMPC) strategy for the BioPower 5 CHP plant that incorporates a detailed non-linear oxygen-balance sub-model. Fast disturbance rejection is achieved by manipulating the primary-air flow, whereas the secondary-air flow fine-tunes the excess-air ratio. The underlying process model was validated experimentally on a full-scale BioPower facility equipped with BioGrate technology capable of firing biomass with up to 65 wt.% moisture. The NMPC was subsequently evaluated in a hardware-in-the-loop simulation environment. The weighting matrices were taught with PPO reinforment learning to optimize the control performance. This RL-based tuning automates what is traditionally a manual, heuristic process, while constraining all weight changes within ±50 % of their original values to preserve closed-loop stability. The results demonstrate superior tracking accuracy, tighter oxygen control, and enhanced robustness to fuel-moisture variations. These findings are presented and discussed.