Neural modelling of biochemical systems using CDTA with adaptive learning rate

Artificial neural networks are becoming increasingly popular as tools to represent chemical/biochemical process dynamics. A combined design and training algorithm (CDTA) is proposed in this study to simultaneously perform network architecture selection and training of the network weights. The proposed CDTA uses a modified back propagation learning method based on a new adaptive learning rate mechanism. The performance of the CDTA is illustrated through application to two typical fermentation systems, namely a SISO anaerobic digester and a MIMO continuous stirred tank fermenter. The adaptive learning rate method is shown to achieve faster convergence than the fixed learning rate method. The network generated is able to represent the fermentation processes accurately, which is illustrated through the reproduction of the training data set as well as through generalization data sets of responses to combinations of load disturbances and set point changes in both SISO and MIMO cases. This study demonstrates that the proposed CDTA can be used to find a neural network representation of chemical/biochemical process dynamics.