Neural network study on the dynamic identification of a biosystem

The objective of this paper is to propose a backpropagation neural network (BPN) with a saturation-type transfer function x÷1+|x| for the study of dynamic identification and prediction of a bioreaction system. Both the dynamic learning and prediction abilities of this network were thoroughly studied. In this work, neural networks provide a dynamic learning and prediction process that moves along the time schedule batchwise. Several factors affecting the learning and prediction performance were studied. Proper learning factor η and momentum term coefficient α as well as the suitable network size were searched. Three-layered BPNs were used in the tests. Different numbers of hidden nodes were tested for learning ability. The best values of η and α were also determined. Proper number of input/output nodes was also studied. Effect of different dynamic learning intervals, either with different starting points and same ending point or with same starting point and different ending points, both on the learning and prediction performance were studied. Different data sampling intervals were also compared based on the performance of dynamic prediction.