Neural networks applied to the prediction of fed-batch fermentation kinetics of Bacillus thuringiensis

This paper proposes using a new recurrent neural network model (RNNM) to predict and control fed batch fermentations of Bacillus thuringiensis. The control variables are the limiting substrate and the feeding conditions. The multi-input multi-output RNNM proposed has twelve inputs, seven outputs, nineteen neurons in the hidden layer, and global and local feedbacks. The weight update learning algorithm designed is a version of the well known backpropagation through time algorithm directed to the RNNM learning. The error approximation for the last epoch of learning is 2% and the total learning time is 51 epochs, where the size of an epoch is 162 iterations. The RNNM generalization was carried out reproducing a B. thuringiensis fermentation not included in the learning process. It attains an error approximation of 1.8%.