Artificial Neural Network
Industrial and
Engineering Chemistry Research, 42(8), 2003, 1707-1712.
Laurentiu A. Tarca, Bernard P.A. Grandjean and Faїçal Larachi
Department of Chemical Engineering & CERPIC
Laval University - Québec, Canada G1K 7P4
To increase confidence in neural
network modeling of multiphase reactor characteristics, we have to take
advantage of some a priori knowledge of
the physical laws governing these systems in order to build neural models
having phenomenological consistency
(PC). A common form of PC is the monotonicity
constraint of a characteristic to be modeled with respect to some important
dimensional variables describing the multiphase system. When the inputs of a
neural model are functions (usually dimensionless) of the variables with
respect to which monotonicity is expected, the monotonicity might not be guarantied, but such a drawback
is only observed after the training. A genetic algorithm based methodology was
proposed to produce several highly accurate and nearly phenomenologically
consistent networks differing by their inputs and architecture. PC and accuracy
were shown to be boosted up meaningfully by combining such networks in a linear
meta-model. A new optimality
criterion for the meta-model parameters identification was proposed and the
results were compared with classical MSE optimality criterion. The proof of
concept of the approach was illustrated in modeling the two-phase pressure drop
in counter-currently operated randomly packed beds.
Keywords
neural network, meta-model, phenomenological consistency, pressure drop,
counter-current packed bed