In this study a data-driven soft-sensor has been developed
for the on-line estimation of ethanol concentrations in batch
distillations plants, at both laboratory and industrial scales.
The soft-sensors applied in wine stills provided good prediction
accuracy of distillate ethanol concentration and even
better performance can be achieved using data pre-processing
procedures. From lab-scale validations, prediction
errors as little as ±0.6% v/v were observed. The soft-sensors
proved successful for industrial distillations too, although
prediction errors were higher (±1.6% v/v) since a less accurate
reference method was used. An important additional
advantage of the soft-sensor proposed here is that it can
reduce delays in ethanol measurements, which should in turn
improve automatic control loop performance.
For this application, and presumably also in other small
to medium scale distillation columns involving a reduced
number of separation stages, our results showed that static
neural network composition estimators have equivalent or
marginally better performance than dynamic estimators.
Even though neural networks are black-box models, static
estimators implicitly represent mixture thermodynamics
and are less dependent on the dynamics of any particular
distillation equipment. In essence this means static estima