Neural Modeling of the Selective Noncatalytic Reduction Of Nitric-Oxide in a laboratory Reactor And A Full-Scale Plant

Neural networks have been used for estimating multivariate models of the selective non-catalytic reduction (SNR) of NO with NH3. A neural network model accurately fitted the data from well-defined laboratory experiments; this model predicted NH3 and NO(x) concentrations at the reactor outlet as functions of a wide variety of parameters. The control scheme of a commercial SNR process was optimised by a different neural model, designed from experimental measurements of NH3 and NO(x) emissions as functions of various operational parameters- Discrete prediction of NH3 was necessary because of limited analytical accuracy. The results demonstrate the power of neural networks in the static input/output modelling of non-linear processes, either generating an accurate model from an accurate set of data or extracting an adequate model from a noisy set of data.