A Multi-Criteria Decision Support Framework for Agricultural Fuel Demand Forecasting: Comparative Analysis of Machine Learning Approaches with Domain-Specific Feature Engineering

This study presents a comprehensive multi-criteria decision support framework for agricultural fuel demand prediction using machine learning algorithms. The research utilizes real operational data from Tarnet A.S., comprising 21,187 daily transaction records collected from agricultural cooperatives in Antalya, Türkiye, spanning from 2019 to 2021. A novel Agricultural Fuel Demand Feature Taxonomy (AFDFT) is proposed, organizing 844 features into five hierarchical categories: temporal, agricultural activity, operational, environmental, and economic factors. Eight machine learning algorithms were systematically evaluated using a multi-criteria approach that considers accuracy, statistical significance, effect size, and performance stability. The XGBoost algorithm emerged as the optimal solution, achieving 78.01% classification accuracy with statistically significant superiority (Friedman test χ²(7)= 23.15, P