COMPARISON OF PREDICTION PERFORMANCES BETWEEN MODELS OBTAINED BY THE GROUP METHOD OF DATA HANDLING AND NEURAL NETWORKS FOR THE ALCOHOLIC FERMENTATION RATE IN ENOLOGY

The modelling of winemaking processes, to predict as far ahead as possible the fermentation performance, is necessary for enhanced supervision and to enable appropriate corrective action to be taken to remedy incorrect fermentation before it is too late. In this paper, we briefly present two heuristic modelling methods-the Group Method of Data Handling (GMDH) and Neural Networks (NN)-which can be used to obtain unstructured models. The identification and prediction performances of the models obtained with these two methods are compared with respect to the alcoholic fermentation rate (dCO2/dt) at five prediction horizons and for four fermentations. It is shown that predictive models obtained with neural network methodology are more accurate than those obtained with GMDH. On the other hand, GMDH models are more versatile when used for the prediction of the fermentation rate of a different fermentation than the one used in the learning process.