Previsão das propriedades do biodiesel e seu perfil ideal de ácidos graxos por meio de aprendizado de máquina

Abstract

Biodiesel is a fuel produced from renewable resources, such as vegetable oils and animal fats. It is considered a cleaner and more sustainable alternative to conventional diesel, which is produced from petroleum. Biodiesel can be defined in terms of ester mixtures as a fuel composed of monoalkyl esters of long-chain fatty acids, which are generally produced by a transesterification reaction. This dissertation addresses the prediction of biodiesel properties and its ideal fatty acid profile through machine learning. The properties of interest are density, kinematic viscosity, and heat of combustion. The study aimed to create models that predict the properties of biodiesel based on its fatty acid composition, allowing the identification of the best compositions for the biofuel in question. The model employed was the k-nearest neighbors (KNN), a classic regressor within machine learning. The methodology details the database used for model creation, which was consulted in the literature and expanded via interpolation, computing 136 samples and their composition in terms of esters derived from acids with 8 to 18 carbons in the chain, containing unsaturations or not. Basic data analysis was performed to compare the original database and the expanded database. The KNN model was accurate in performing the regression of the 3 properties of interest. Interpretability methods were used to check its adherence to the physical phenomenon, namely: importance scales and partial dependence plots. It was observed that the model is capable of reasonably representing the expected effects in terms of ester composition: mixtures rich in short-chain saturated fatty acids are less viscous, denser, and have lower heat of combustion. The differential evolution method allowed the search for mixture compositions that yield the highest possible values of heat of combustion, around 40 MJ/kg.