Abstract: In response to the problems of strong combustion nonlinearity, tight coupling of multiple variables, and poor adaptability of traditional "constant oxygen" control strategies in deep peak shaving of coal-fired units under the "dual carbon" target, a multi-dimensional information fusion based coal-fired boiler wide load combustion optimization layered control method is proposed. Firstly, a three-layer collaborative architecture of "perception decision execution" is constructed, which integrates multidimensional information such as SCR inlet NO_<i>x, tail flue CO concentration, and water wall temperature, breaking the limitation of traditional control relying only on a single variable of load; secondly, establish a model predictive control (MPC) strategy based on recursive least squares (RLS) online identification, and use the principle of thermal balance to construct coal quality soft sensing indicators as feedforward compensation. This method explicitly addresses the hard constraints of wall temperature safety and environmental protection in the objective function, achieving dynamic cruising of oxygen set values and synergistic optimization of secondary air/SOFA air from a mechanistic perspective. Theoretical analysis shows that this layered control strategy can effectively solve the contradiction between low load stable combustion and emission control, providing a new engineering paradigm for improving the economy and safety of unit operation under complex working conditions.