Dynamic hybrid modeling of LSTM-boosted mechanism and adversarial generation for industrial fuel ethanol fermentation process

Abstract

Fuel ethanol has great potential as a substitute for non-renewable energy. However, modeling the pragmatic industrial process of fuel ethanol fermentation continues to be an exceedingly formidable challenge. First, the kinetic parameters in pure mechanism modeling ignore the significant differences among batches. Second, the prediction of ethanol and sugar concentrations is highly dependent on the yeast concentration, which is often missing due to factory technology. Thus, we propose a dynamic hybrid model to realize the accurate prediction of ethanol and sugar concentrations. First, the kinetic parameter extractor based on long short-term memory models patterns of kinetic parameters and performs as a generator. Second, yeast experiments are conducted in the laboratory, and an adversarial generation of the yeast module is designed to realize yeast concentration generation, forcing the distribution of generated yeast concentrations near the distribution of laboratory ones. Finally, our proposed differential equation module decoder supplements vital mechanisms. Experimental results show that our model is biologically significant with low prediction errors.