Abstract: The carbon-neutral policy puts forward higher requirement for the load peak shaving capacity of coal-fired power plants. However, the NO_x concentration in the flue gas will fluctuate greatly while the load of the unit changes in a large range, which makes it more difficult to control ultra-low NO_x emission. Aiming at the problem that it is difficult to quickly and accurately control the spray amount of ammonia under large-scale variable load conditions, this paper took the combined denitrification system of a circulating fluidized bed power plant as the research object, and established a control strategy that coupled key parameter feedforward correction and multi-model predictive control. Using the furnace outlet flue gas temperature as the basis to divide the working condition sub-models, the parameters of each sub-model were identified according to the step disturbance tests and the improved particle swarm optimization. Finally, the multi-model controller was established through the degree of membership weighted method. Engineering application result shows that, the feedforward modified multi-model predictive control method has a fluctuation range of ±5.8mg/m³ under steady load conditions, and it is ±8.1mg/m³ under variable load conditions. The standard deviations are 2.10mg/m³ and 2.89mg/m³ respectively. The application result proves the effectiveness of this control method.