Answers are grounded in this paper and cite their sources.
Reliable phosphate soft sensing requires predictive accuracy to be evaluated alongside data leakage, input availability and the operational burden of surveillance. Using two years of full-scale supervisory control and data acquisition records, hourly reactor-phosphate nowcasting was assessed with twelve machine-learning regressors, feature-availability ablation, repeated model fits, rolling-origin evaluation and season-stratified predictions, with expert fitting, gate fitting and residual calibration assigned to separate chronological blocks. A no-dose configuration, which retained ammonium but excluded all current and historical metal-dose inputs, achieved a March–July 2023 test R 2 of 0.470 ± 0.005 across ten seeds, versus 0.697 ± 0.006 for the control-inclusive configuration, indicating that much of the apparent accuracy depends on controller-response inputs; across five rolling-origin windows, the corresponding mean R 2 values were 0.398 and 0.670. Random rather than chronological splitting inflated test R 2 from 0.688 to 0.834, showing how temporal leakage overstates accuracy. A one-sided 90% upper-bound rule with a 6-h cooldown detected 43 of 73 excursions, with 3.21 false notifications per week. Symmetric 90% intervals covered 91.4% of observations overall but only 29.2% of high-phosphate states. The workflow provides a quantitative basis for evaluating auxiliary reactor surveillance and prioritising prospective tests with independent reference measurements, analyser outages and controller changes.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: