Machine Learning-Based Prediction of TransportRelated CO₂Emissions Across Heterogeneous National Contexts

Jawher Bouaziz, Najah Kechiche, Khdeijetna Mohamed El Moctar, Walid Hassen

In this paper, we investigate transport-related CO₂emissions (TCO2) across 31 countries using a set of socioeconomic, energy, sectoral, and meteorological factors as well as temporal indicators. Correlation analysis reveals a large variability between the studied countries, with socioeconomic indicators and sectors showing consistently positive and negative relationships with TCO2 emissions. In contrast, renewable energy consumption generally exhibits negative correlations. To estimate TCO2, we used three machine-learning models, such as Support Vector Regression (SVR), Ridge Regression (RR), and Random Forest (RF). The best model was RF (R² = 0.87, MAPE = 3.08%), followed by SVR with moderate accuracy, while the performance of RR was inadequate to predict the TCO2. In terms of national-level predictions, the RF method has good generalization ability, with very high accuracies (<10% rRMSE) obtained in several developed nations (i.e., Germany, Denmark, the United States, France, and Portugal) along with reasonable performance in developing countries (i.e., China, India, and Tunisia, etc.). However, the regression plot shows a strong agreement between observed and estimated values, with only minor deviations, particularly at milder and higher emission levels. These results confirm that the RF approach can be applicable in heterogeneous contexts.