Methods for Parameter Estimation in Wine Fermentation Models

The estimation of parameters in a wine fermentation model provides the opportunity to predict the rate and concentration outcomes, to strategically intervene to change the conditions, and to forecast the rates of heat and carbon dioxide release. The chosen parameters of the fermentation model are the initial assimilable nitrogen concentration and yeast properties (lag time, viability constant, and specific maintenance rate). This work evaluates six methods for parameter estimation: Bard, Bayesian Optimization, Particle Swarm Optimization, Differential Evolution, Genetic Evolution, and a modified Direct Grid Search technique. The benefits and drawbacks of the parameter computational methods are discussed, as well as a comparison of numerical integration methods (Euler, Runge-Kutta, backward differential formula (BDF), and Adams/BDF). A test set of density-time data for five white and five red commercial wine fermentations across vintage, grape cultivar, fermentation temperature, inoculated yeast strain, and fermentor size was used to evaluate the parameter estimation methods. A Canonical Variate Analysis shows that the estimation methods are not significantly different from each other while, in the parameter space, each of the fermentations were significantly different from each other.