Aiming at the problem that it is difficult to accurately and timely measure the inlet NOx concentration in the denitrification system of selective catalytic reduction (SCR) in thermal power plants, due to the excessive factors affecting the inlet NO_x concentration and the large delay and inertia of the system, the Max-Relevance and Min-Redundancy (mRMR) combined with Bayesian optimization (BO) algorithm is proposed, optimize the dynamic soft measurement model of NO_x concentration at the inlet of the SCR denitration system of the stacking ensemble model. Aiming at the problem of reduced prediction accuracy of static single model and asynchronous timing of auxiliary variables and inlet NOx concentration in the process of dynamic NO_x generation, the mRMR-BO combined with model was used to screen the auxiliary variables, Copula Entropy (CE) determined the delay of auxiliary variables, the BO combined with model determined the order of auxiliary variables, and TCN and LASSO were integrated by Stacking method. The auxiliary variables containing delay time and order information were used to construct a dynamic stacking ensemble soft measurement model, and the simulation results showed that the root mean square error, average absolute error, and average absolute percentage error of the integrated model compared with TCN and LASSO single networks were the smallest. Compared with the static ensemble model, the dynamic ensemble model has higher prediction accuracy and can achieve accurate soft measurement of the inlet NO_x concentration.