Fermentação alcoólica: desenvolvimento de metodologia para o cálculo de eficiência e modelagem por redes neurais de unidade de fermentação industrial

Abstract

The search for alternative sources of fuel, economically and environmentally viable, that could replace fossil fuels has increased in the last years. Brazil occupies a prominent position regarding the pioneering development and ethanol production from sugarcane in large scale. Despite the consolidated technology studies have shown that there is room for improvement and innovation on the industrial ethanol production process, through the development of new production strategies. Thus, the objective of the present work is justified. In the first step, it was developed a new methodology based on mass balances to calculate the ethanol fermentation efficiency of the process operated in fed-batch mode. The new methodology was compared to others that have already been used by the industrial sector. Data from fed-batch fermentations in bench and industrial scales were used. The new methodology allowed assessing with more precision the ethanol production of the industrial unit. It was also possible to calculate the quantity of overestimated ethanol when fermentation efficiency was calculated by the established methodologies. The second step of this work was to develop a model based on artificial neural networks for representing the fermentation process. Data of an industrial fermentation unit were used. The objective was to assess the effects of different industrial variables on ethanol production. Important input variables that could influence the ethanol fermentation efficiency were identified. Then, it was set up an artificial neural network to estimate the final ethanol concentration in the industrial bioreactors. After the network was trained, it was used together with a stochastic optimization algorithm based on population (Particle Swarm Optimization – PSO) to estimate the increase of the ethanol concentration in the fermentation process by seeking optimum values for the input variables. The training, validation and test steps of the artificial neural network were performed using 200 points of the industrial data. After the update of the weights has been completed during the training step, the neural network model was tested using the validation data set and it predicted the ethanol concentration in the process reaching 0.91 for the determination coefficient (R²) and a mean squared error of 0.26. The model was tested with new experimental data after the training, and it was obtained relative deviations below 4.0%. This fact illustrates the prediction potential of the artificial neural network model. In the optimization step of the input variables, it was possible to reach an increase of 1.0 °GL in the ethanol concentration at the end of the process. Therefore, there is a possibility to use this tool on the industrial process, aiming to increase the industrial ethanol production.