Modeling flash points of biofuels using neural networks

Abstract

The search for renewable energy resources is driven by environmental hazards caused by petroleum derivatives, price fluctuations, and the unsustainability of fossil fuels. In Brazil, biodiesel and bioethanol are established renewable fuels, while butanol shows promise as an alternative fuel, requiring research into their safety and efficiency. The Flash Point (FP) is crucial for flammability assessment and safety in combustion processes, but its experimental measurement is resource-intensive. This study evaluates the capability of artificial neural networks (ANNs) to predict FP for some biofuels and their blends, using a dataset of 490 points. Notably, 24 of these points were newly acquired, while the remaining 466 were sourced from literature. A robust ANN model was trained using a 5-fold cross-validation with an 80/20 data split, incorporating average molar mass, vapor pressure natural logarithmic, and experimental method as input features. The final model, featuring three hidden layers determined through a parametric analysis, achieved a Root Mean Square Error (RMSE) of 4.22 K and a Mean Absolute Error (MAE) of 3.09 K for 98 unknown points. The model achieved satisfactory accuracy, with MAE ranging from 1.51 K to 3.63 K, and performed comparably to traditional UNIFAC thermodynamic models. These results highlight the potential of ANNs for FP prediction across diverse datasets.