Research on adaptive closed-loop combustion optimization control method of boiler under dynamic working condition based on mechanism modeling and deep learning learning

In response to the strong nonlinearity, multiparameter coupling and dynamic operating conditions fluctuations in industrial combustion systems, the traditional single mechanism model has insufficient accuracy due to simplified assumptions, and the pure data-driven model lacks physical interpretability, making it difficult to achieve highprecision closed-loop control. This paper proposes a closed-loop combustion optimization control method that integrates mechanism modeling and deep learning: build physical constraint frameworks such as energy conservation and chemical reaction balance of the combustion process through mechanism models to ensure the reliability of the model basics; use deep learning models to capture the correlation between complex dynamic characteristics and implicit parameters, and improve the adaptability of multiple conditions. A real-time acquisition system including furnace temperature, flue gas composition, fuel/air flow rate and other parameters was studied and built. The fusion model was trained based on the constraints derived from 7200+ hours of on-site operation data and mechanism, and a closed-loop control architecture of “perception-prediction-decision-executionfeedback” was designed, and real-time optimization was achieved in combination with adaptive model prediction control. Experimental results show that compared with traditional PID control and pure deep learning control, the combustion efficiency of this method is increased by $0.95 \%$ under stable operating conditions, and the emission of nitrogen oxides ($\mathrm{NO}_{\mathrm{x}}$) is reduced by $10.3 \%$. The research verifies the advantages of the fusion model in taking into account both physical interpretability and dynamic adaptability, and provides a feasible solution for efficient lowemission closed-loop control of complex combustion systems.