Empirical modeling of ethanol production dynamics using long short-term memory recurrent neural networks

• Published in: Bioresource technology reports 2021-09, Vol.15, p.100724, Article 100724
• Main Authors: Sousa, Felipe Matheus Mota, Fonseca, Rodolpho Rodrigues, da Silva, Flávio Vasconcelos
• Format: Article
• Language: English
• Subjects: Deep learning; Empirical modeling; Fermentation; Long short-term memory; Recurrent neural networks
• ISSN: 2589-014X; 2589-014X
• DOI: 10.1016/j.biteb.2021.100724

Long short-term memory networks (LSTM) were trained to predict the dynamics of a fermentation process with varying kinetic parameters. The training and test database was obtained through simulations using phenomenological equations and an artificial neural network (ANN) to adjust the kinetics according to operational conditions. Results showed that a shallow LSTM was able to predict the dynamics of all endogenous variables accurately using only three timesteps, including the inverse responses observed, a difficult feature to incorporate in empirical models. For tests using the predictions recursively as endogenous input variables, deeper structures were necessary to achieve a reasonably good performance with errors within commercial sensor's accuracy. These results indicate the applicability of LSTMs to model fermentation processes using raw data, abstractly incorporating kinetics variations in the model, and suggest its use as a tool in model predictive controllers or for process optimization. [Display omitted] •Shallow LSTM modelled the dynamics of a fermentation process with varying kinetics.•Shallow LSTM was unable to predict using the endogenous variables recursively.•Deeper LSTM enhanced the recursive performance with errors within sensor's accuracy.