Risk-controlled minimization of flue-gas abatement reagents in iron-ore sintering: A causal, adaptive machine-learning framework for integrated pollution control

초록

Iron-ore sintering is a major source of nitrogen oxides (NOx) and sulfur oxides (SOx), and abating them consumes large amounts of denitrification (ammonia) and desulfurization reagents, which raises cost and creates secondary pollution such as ammonia slip and spent absorbent. Minimizing reagents while safely meeting emission limits is therefore a problem of integrated pollution control and process safety. It is hard for three reasons: dosing responds to emissions, so the effect is confounded; the process drifts; and predictions carry error. Using integrated operating, abatement-facility, and raw-material data from a real sinter plant, we propose a single pipeline linking causal inference, adaptive prediction, and risk-controlled optimization. Double machine learning corrects the desulfurizer effect from a naive + 0.34 to a causal −0.10, confirming that more reagent lowers emission. Under leakage-free, time-ordered evaluation, periodic retraining recovers predictive power (R2: SOx 0.50, NOx 0.44). A safety-margin-aware optimization cuts reagents by about 16% (desulfurizer) and 19% (ammonia) at a controlled 10–20% exceedance risk, reducing secondary pollution without cross-media transfer. A NOx–SOx asymmetry emerges: NOx is controllable (early-warning AUC 0.82), whereas SOx control saturates near 84% coverage because of a daily-resolution observability limit (the residual is unexplained by measurements yet persistent). © 2026 Institution of Chemical Engineers. Published by Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

키워드