Prediction of carbon emissions from vehicles using interpretable neural networks

Abstract

Vehicle carbon dioxide emissions are an important environmental concern that requires prediction methods with both high accuracy and transparent explanations. This study proposes an interpretable neural network (INN) framework for estimating vehicle carbon dioxide emissions using a public Canadian fuel-consumption and emissions dataset. A feed-forward neural network is employed for regression, while SHapley additive exPlanations (SHAP) are used to interpret the contribution of each input feature. The proposed model is compared with linear regression, Bayesian regression, support vector regression (SVR), and extreme gradient boosting (XGBoost). Experimental results show that the neural-network model achieves a root mean squared error (RMSE) of 4.98, a mean absolute percentage error (MAPE) of 0.0128, and a coefficient of determination (R-squared) of 0.9926, outperforming all baseline models. SHAP analysis identifies combined fuel economy, city fuel consumption, and highway fuel consumption as the most influential predictors. Higher fuel economy is associated with lower predicted emissions, whereas higher city and highway fuel consumption increases predicted emissions. These findings demonstrate that the proposed framework provides accurate and interpretable vehicle carbon dioxide emission estimation.