One-step-ahead Product Predictor for Profit Optimization of Penicillin Fermentation

In antibiotics industry, the titre in bioreactors is the most important process variable both for process supervision and scheduling. It is therefore of great significance to develop a software sensor to predict the product formation. In this contribution, a pseudo dynamic product predictor based on artificial neural network is designed. The input process variables of the predictor include substrates, precursor, nutrients and oxygen consumption, carbon dioxide evolution and also the product formation. These process variables are usually available in practice. Only the accumulated values of these variables rather than the instant ones are used. A feedforward neural network is chosen for one-step-ahead prediction of total product titre at the next step. The input vector of the neural network is the time series of process variables over a predetermined time period. The database for network training is composed of a series of such vectors taken from historical charges which are processed as well as updated by a moving-data-window’s technique. The software sensor is tested by data of three industrial charges and ten model "generated" batches. MATLAB is used for training and testing of the neural network. The predictor is expected to be applied in optimal fermentation time scheduling for antibiotics production. Improvements and further development of the software sensor for real time applications are discussed.