Implementação de algoritmos de aprendizado de máquina para predição da produção de biogás em escala industrial

Abstract

One of the key players in the transition to more sustainable energy sources, biogas faces the challenge of biomass source variability in its industrial operation, significantly impacting the predictability of its production. This unpredictability complicates operational planning and process optimization and ensures consistent and efficient production. Therefore, this work aims to develop a machine learning algorithm to predict biogas characteristics in an industrial plant, using data from raw materials and operational conditions to estimate biogas generation and the methane fraction contained within it. Initially, by organizing available datasets and statistical and exploratory analyses, it was possible to identify trends and correlations between system parameters and outcomes. Subsequently, by applying modeling and machine learning techniques, it was possible to develop and validate predictive algorithms, including Random Forest (FR), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (RNA). The highest prediction efficiency was obtained by XGBoost, followed by FR and RNA, with R² values of 0.74, 0.71, and 0.70, respectively, with XGBoost achieving an average R² of 0.82 in the scenario considering both raw material and operational characteristics. This scenario proved to be more efficient due to the increased robustness provided by more variables.