RESUMO
Machine learning algorithms provide detailed description of the anaerobic digestion process, but the impact of data preparation procedures and hyperparameter optimization has rarely been investigated. A genetic algorithm was developed for optimizing data preparation and model hyperparameters to simulate dynamic methane production from steady-state anaerobic digestion of agricultural residues at full-scale. A long short-term memory neural network was used as prediction model. Results indicate that batch size, learning rate and number of neurons are the most important model parameters for accurate description of methane production rates, whereas combination of hyperparameter and data preparation optimization shows best model efficiencies, with a root mean square scaled error of 76.5 %. Mass of solid feed, time and mass of volatile solids are the most relevant input features. This study provides fundamental steps for optimal prediction of dynamic biomethane production, as a reliable basis for improving bioconversion efficiency during anaerobic digestion of agricultural residues.
