Multi-indicator Prediction and Multi-Target Model Predictive Control for Coordinated Regulation of NOx Emissions and Ammonia Slip in Cement Kiln Denitrification

Abstract

Accurate regulation of nitrogen oxide (NOx) emissions is essential for stable and environmentally compliant cement production. Cement kiln flue gas denitrification is characterized by strong nonlinearity, long transport delays, multivariable coupling, and uncertain operating disturbances, which make simultaneous NOx control and ammonia-slip suppression difficult. This study proposes a multi-indicator prediction and control framework for cement kiln denitrification. First, a multi-indicator prediction model, named MI-ILTA, is developed by integrating an improved long short-term memory network with a multi-kernel temporal convolutional network and an attention-based feature weighting module. The model is designed for multi-step prediction of NOx concentration and ammonia slip. Based on the predicted outputs, a weighted-sum scalarized Multi-Target Model Predictive Control (MT-MPC) strategy is established to coordinate NOx set-point tracking, ammonia-slip suppression, and smooth ammonia-flow adjustments under process constraints. A differential evolution algorithm solves the rolling optimization problem and determines the control inputs. Industrial data from a cement-plant distributed control system are used to compare the proposed model with TCN, LSTM-TCN, and LSTM-TCN-Attention baselines. The MI-ILTA model achieves lower MAE, RMSE, and SMAPE across the reported prediction horizons for both NOx concentration and ammonia slip. In industrial-data-based closed-loop simulations, the proposed MT-MPC strategy reduces the settling time from 552 s to 366 s and overshoot from 4.1% to 0.38% compared with PID control, while maintaining lower ammonia slip. These results indicate that the proposed framework can improve dynamic regulation and emission stability in cement kiln denitrification systems.