Physics-inspired perspective on synergistic optimization: a deep receding-horizon optimization strategy for...

To address the multi-objective conflicts among nitrogen oxides (NO_x) emission compliance, ammonia slip constraints, and environmental cost control in the denitrification process of municipal solid waste incineration (MSWI), a physics-inspired intelligent ammonia injection optimization method based on a deep learning prediction and receding-horizon optimization (dl-RHO) framework is proposed. A synergistic prediction model incorporating time derivative (TD) constraints, termed Conv-Transformer-TD, was developed. The convolutional neural networks and Transformer are utilized for deep temporal feature extraction while an innovative TD loss is introduced to eliminate the phase lag inherent in traditional predictions. Furthermore, NO_x feature is explicitly incorporated to assist NH₃ prediction, thereby characterizing chemical reaction equilibrium. Experimental results demonstrate that the model achieves coefficients of determination (R²) of 0.90 and 0.94 for NO_x and NH₃ predictions on the test set, with root mean square errors of 3.70 mg/m³ and 0.31 mg/m³, respectively, significantly outperforming common benchmark models. Ablation experiments confirm the critical roles of each module within the model. Building upon this, a closed-loop RHO framework based on particle swarm optimization was constructed to achieve multi-objective rolling horizon optimization. In a 144-hour robustness verification case under novel conditions, the dl-RHO strategy, compared to actual manual operation, reduced the average NO_x concentration by 22.09 % and the mean ammonia slip by 40.35 %, while achieving a cumulative saving of 19.84 % in ammonia water consumption, all while ensuring NO_x emission compliance. The multi-objective synergistic optimization problem in large-lag systems is successfully addressed, providing an efficient engineering solution for the intelligent operation of MSWI processes.