In
this study, the performance of three black-box identification
techniques using linear autoregressive with exogenous input (ARX),
nonlinear ARX (NARX), and Hammerstein–Wiener (HW) algorithm
to model the dynamics of UV/H2O2 continuous
tubular photochemical reactor for the treatment of poly(vinyl alcohol)
(PVA) based on experimental data is investigated. In addition, the
inherent nonlinearity of the reaction process is assessed. The reactor
dynamics in the NARX model is estimated by wavelet, sigmoid, and tree
partition networks along with the assessment of the performance of
each model. Although a sigmoid network describes the nature of chemical
processes better, the results show that tree partition network-based
NARX is the most suitable estimator for the studied process as represented
by its highest quality of fit (91.59% for training data set and 88.17%
for validation of data set), lowest loss function (mean-squared error,
MSE) (0.0004279), model realizability, open-loop stability, model
whiteness, and model independence.