Abstract:Abstract: To address the modeling difficulties of thermal power unit combustion systems under wide-load operating conditions, which are caused by complex mechanisms, strong coupling, and pronounced nonlinearity, a parallel hybrid modeling method based on mechanism-data gated fusion is proposed. First, a mechanism-data-driven parallel hybrid modeling framework is constructed, in which a simplified mechanism model and a data-driven model are used as two parallel prediction branches, and their outputs are adaptively weighted and fused through a gated fusion network. Specifically, the simplified mechanism model is used to represent the main steady-state relationships and dynamic operating mechanisms of the system, thereby ensuring model generalization. The data-driven model, which combines a convolutional neural network and a long short-term memory network, is used to extract multivariate time-series features and learn the nonlinear characteristics that are difficult for the mechanism model to describe accurately. In addition, an error feedback mechanism is introduced to dynamically correct prediction deviations and improve the prediction accuracy of the model under varying operating conditions. The gated fusion network learns the mapping relationship from operating features to fusion weights, fully exploiting the complementary advantages of the mechanism model and the data-driven model under wide-load varying conditions, and enabling the fusion weights to be adaptively adjusted with operating conditions. Finally, simulation validation is conducted on the combustion system of a 350 MW circulating fluidized bed boiler. The simulation results show that the proposed method outperforms the comparison models in terms of prediction accuracy and dynamic response performance.