Predictive modelling of brewing fermentation: from knowledge-based to black-box models

Abstract

Advanced monitoring, fault detection, automatic control and optimisation of the beer fermentation process require on-line prediction and off-line simulation of key variables. Three dynamic models for the beer fermentation process are proposed and validated in laboratory scale: a model based on biological knowledge of the fermentation process, an empirical model based on the shape of the experimental curves and a black-box model based on an artificial neural network. The models take into account the fermentation temperature, the top pressure and the initial yeast concentration, and predict the wort density, the residual sugar concentration, the ethanol concentration, and the released CO 2. The models were compared in terms of prediction accuracy, robustness and generalisation ability (interpolation and extrapolation), reliability of parameter identification and interpretation of the parameter values. Not surprisingly, in the case of a relatively limited experimental data (10 experiments in various operating conditions), models that include more process knowledge appear equally accurate but more reliable than the neural network. The achieved prediction accuracy was 5% for the released CO 2 volume, 10% for the density and the ethanol concentration and 20% for the residual sugar concentration.