Cephalosporin C fed-batch cultivation undergoes great fluctuations. Some key state variables, such as product concentration and carbon source consumption, are very difficult to measure on-line, while these variables are essential to process monitoring and control.A neural network based software prediction of the key state variables for cephalosporin C fed-batch fermentation was investigated. A rolling learning-prediction procedure was used to deal with the time variant property of the process, and was also demonstrated to be beneficial to improving prediction accuracy.The successful prediction of the product formation enabled on-line evaluation of the economic performance of a charge and made optimal scheduling possible.The prediction approach was validated with the data of 49 industrial charges.
