Abstract

Batch distillation is a traditional and widely-used technique to produce Pisco brandy, a young spirit made from Muscat wine. It is necessary to track a given ethanol composition in the distillate in order to obtain a reproducible spirit with a desired aromatic profile. The use of multiple ethanol sensors represents a considerable cost, which prevents many distilleries from adopting this technology. Aiming to provide practical and affordable industrial-scale distillation control technology, we developed a soft-sensor to estimate distillate ethanol concentration on-line based on four temperature measurements in the still. The soft-sensor, calibrated with laboratory and industrial experimental data, consisted of an Artificial Neural Network and involved simple data pre-processing procedures. Simplicity and good performance were the metrics adopted for testing different algorithms and network structures. Returning mean prediction errors of ±0.6% v/v with laboratory scale distillations and ±1.6% v/v in industrial trials, the resulting accuracy of the soft-sensor is sufficient to improve standard practice and reproducibility. © 2008 Elsevier Ltd. All rights reserved.