Data-driven prediction of the product formation in industrial 2-keto-l-gulonic acid fermentation

Mixed culture fermentation of Bacillus megaterium and Gluconobacter oxydans is widely used to produce 2-keto-l-gulonic acid (2-KGA), a key precursor for l-ascorbic acid synthesis. For such mixed cultivation, kinetic modelling is difficult because the interactions between the two strains are not well known yet. In this paper, data-driven prediction of the product formation is presented for the purpose of better process monitoring. A rolling learning-prediction approach based on neural networks is practiced to predict 2-KGA formation. Techniques associated with the approach, such as the data pretreatment and the rolling learning-prediction mechanism, are given in more detail. The validation results by using the data from commercial scale 2-KGA cultivation indicate that the prediction error is less than 5% in the later phase of fermentation and the reliable prediction time span is 8 h. The robustness of the prediction approach is further tested by adding extra noises to the process variables.

► This work demonstrates originally that the ANN-based RLP method may be successfully used to predict the 2-KGA formation. ► The ANN input variables, total product formation and total alkali consumption, are generated by data preprocessing. ► The ANN database undergoes rolling update as the cultivation of the current batch proceeds. ► The ANN structural parameters and the optimal input window width of the RLP method are identified for 2-KGA fermentation by examining the prediction error and the robustness of the prediction approach.