A rolling learning-prediction. approach based on neural networks is propose
d with the aim of on-line prediction of the product formation. Commercial-s
cale penicillin cultivations were taken as an example to test the product p
redictor. Raw data are pretreated in such a way that each input vector of t
he neural network consists of a series of time-discretised values on a spec
ified transient of process variables. The output vector is composed of the
amount of product at the next one and two prediction steps. The process var
iables involved in the predictor include carbon dioxide and product formati
on as well as oxygen, precursor and substrate consumption. Accumulated rath
er than instant values of these variables were used. A simple three-layer f
eedforward backpropagation neural network with a tangent sigmoidal transfer
function in the hidden nodes and a linear one in the output nodes was used
as the main frame of the product predictor. The proposed prediction proced
ure is called rolling learning-prediction because the training database is
updated after each sampling interval and the learning-prediction is repeate
d thereafter. The robustness of the predictor was illustrated by its adapti
ve ability to widely scattered data sets and extra added noises. The testin
g results indicated that a prediction accuracy of 2-5% could be generally e
xpected in the later phase of cultivation and reliable prediction time span
s may take more than 10% of the cultivation period for penicillin productio
n. An intrinsic problem of using neural networks-occasional trap of the net
work in bad local minima-is automatically detected and remedied. In additio
n, it was illustrated by example that the prediction error signal may be po
tentially used to detect extraordinary charges caused, for example, by cont
amination. Problems associated with the industrial application of the predi
ctor are discussed. (C) 1999 Elsevier Science B.V. All rights reserved.