Abstract: To overcome limitations of traditional single-phase coal combustion boiler models, a comprehensive full-process model for a coal-ammonia co-firing boiler was established. An improved affinity propagation-least squares support vector machine (AP-LS-SVM) dynamic multi-modal approach was proposed and applied for model validation and accuracy enhancement. The model consisted of three main modules: fuel pretreatment, combustion chamber dynamics, and key parameter output, which enabled dynamic simulation from fuel properties to boiler outputs. Based on actual operational data from a 600 MW co-firing boiler, the mutual information (MI) algorithm was employed for feature selection and data reconstruction to identify the optimal time delays between input and output parameters, and the proposed AP-LS-SVM method was then used for validation. Results show that the model achieves high goodness-of-fit values, all exceeding 0.950. Compared with conventional methods, prediction errors for all key parameters are significantly reduced, with optimization rates ranging from 22.5% to 58.1%. The method effectively improves the prediction accuracy and adaptability of the model under wide and varying operating conditions with ammonia co-firing ratios from 0% to 30%.