A novel hybrid machine learning model proposal for biodiesel consumption: A feature engineering based predictive framework

In this study, an innovative model based on machine learning is proposed to analyze the factors affecting biodiesel consumption demand and to estimate biodiesel consumption more accurately than the base models. In addition to biodiesel-related metrics, various macro indicators such as hot and cold days, natural gas, and oil substitute energy prices were used for biodiesel consumption estimation in the study. Monthly data between January 2001 and July 2024 were used in the study. New variables were obtained by applying feature engineering to the obtained data set in order to capture historical trends and seasonal fluctuations. In the proposed model, firstly, the feature importance values of the variables used were determined, and then the Ridge, XGBoost, LightGBM and Catboost algorithms were combined with the stacking method to increase the prediction power. The performance of the proposed model was compared with the base models using RMSE, MSE, MAE, and R2 metrics. According to the results, the proposed model gave better results than the base models. Finally, SHAP analysis was performed to evaluate the most important factors affecting biodiesel consumption and the effects of these factors on the model.