softsensor at a different distillation plant. Wine stills at this
plant had a similar rectification column with six trays, but
the reboiler differed: higher load capacity (7%), larger
steam coil and used no thermal insulation. From Table 5
and Fig. 3 the soft-sensor provided reasonable predictions
for ethanol concentration at the second plant and, once
again, the static FFANN estimator performed on average
marginally better than the RANN estimator. Deviations
from real values were more evident at this second plant,
however. The accuracy of both soft-sensors was significantly
lower than at the first plant (compare Table 4 with
Table 5). Large deviations mostly occurred at the beginning
and at the end of the distillation, and 75% of the residuals
lay within ±1.5% v/v, which is sufficient to ensure
good control performance. Nevertheless, more accurate
and generalizable soft-sensors could be developed if data
from different plants or still designs are used for training.