Small-scale waste incinerators commonly employ selective non-catalytic reduction (SNCR) using urea solution. However, effective reagent dosing is complicated by the delay between changes in furnace conditions, urea injection, and the resulting stack NOx measurements. This study developed a physics-guided data-driven framework for forecasting NOx over a control-relevant horizon using operational data from a 48 t/day grate-type waste incinerator.
Operational data collected from June to August 2025 were averaged at 15 s intervals and included stack NOx, boiler-exit O₂, urea-pump frequency, furnace-exit temperature, combustion-load indicators, and waste-feeding variables. Considering an approximately 3 min gas-transport and measurement delay, the forecasting model simultaneously predicted the next 12 time steps using a 15-step input window. Because the furnace-exit thermocouple was embedded in the refractory and therefore exhibited a delayed and smoothed response, a one-dimensional transient heat-transfer model accounting for refractory conduction and gas-side convection and radiation was used to estimate an effective gas temperature. Forecasting performance was compared using either the measured furnace-exit temperature or the physics-derived temperature while keeping the remaining model configuration unchanged.
Cross-correlation analysis showed that the estimated gas temperature led stack NOx by approximately 13 time steps, whereas the measured furnace-exit temperature led it by only about 3 steps. Replacing the measured temperature with the physics-derived estimate reduced the 12-step-ahead RMSE from 5.88 to 5.60 ppm, increased the correlation coefficient from 0.82 to 0.84, and improved R² from 0.67 to 0.70. These results demonstrate that reconstructing an effective gas temperature from the thermocouple response improves temporal alignment with the SNCR control horizon and enhances multistep NOx forecasting. The proposed approach provides a promising basis for predictive urea dosing, although extended field validation is required to confirm its robustness and operational benefits.
 

waste incineration; NOx forecasting; physics-guided learning; SNCR; multi-step prediction