Industrial data was acquired over the course of manual
operation of a wine still under different operating policies.
Data were collected at different times of the year, hence the
composition of the wine as well as environmental conditions
varied considerably. In Fig. 2, ethanol concentration
has been mean-centered and distillation time normalized,
to protect confidentiality of the plant’s operation policies.
In this case, both soft-sensors attained good ethanol concentration
predictions throughout the distillation (Table 4
and Fig. 2). Performances of both FFANN and RANN
were very similar, and most deviations from real measurements
lay within the interval ±1% v/v. Residuals were
somewhat higher toward the end of the distillation in run
2 using the RANN estimator (Fig. 2b), however the mean
absolute error was still only 0.66% v/v (Table 4). As for the
FFANN estimator, the MAE was only 0.61% v/v, which
represents a big improvement over current industrial
practices, in which poor instrument resolution and often
overlooked temperature deviations frequently lead to measurement
errors as high as 6% v/v. Despite these small
errors, predictions are significantly impaired by the low
accuracy of the hydrometer used as reference (±1.0% v/v,
with temperature correction). Thus, the overall accuracy
of the soft-sensor for industrial distillations is estimated
as 1.6% v/v (sum of reference measurement error and
MAE). These results could have been even better had a
more accurate reference measurement method been used,
as the laboratory results demonstrated.