Explainable Artificial Intelligence for National CO2 Emission Prediction and Mitigation: A Multi-Country, SHAP-Based Framework

Abstract

Accurate prediction of national carbon-dioxide (CO₂) emissions is central to evidence-based climate policy, yet the most accurate machine-learning models behave as opaque “black boxes” that give little guidance on which levers actually reduce emissions. This paper presents a reproducible, explainable AI (XAI) framework that couples predictive modelling with transparent driver attribution and a policy-facing counterfactual analysis. Using the open Our World in Data (OWID) CO₂ panel restricted to sovereign nations over 1990–2022 (5,358 country–year records across 164 countries), we engineer socio-economic and energy drivers aligned with the Kaya identity and predict annual emissions under a strictly temporal hold-out that removes look-ahead leakage. Four learners are benchmarked; gradient boosting attains the best generalisation (R² = 0.978 on the log scale, MAPE = 26.5%). The trained model is then interrogated with three complementary XAI methods—SHapley Additive exPlanations (SHAP), permutation importance and partial-dependence profiles—which consistently identify economic scale (GDP) and energy intensity (energy per unit GDP) as the dominant and, crucially, actionable determinants of national emissions. A transparent digital-twin counterfactual for India (2022) shows that a 20% cut in energy intensity lowers predicted emissions by 8.7%, and a combined deep-decarbonisation lever by 17.0%. By making both the model and the recommended levers interpretable on an open dataset with leakage-free validation, the framework turns emissions forecasting from a purely predictive exercise into a decision-support tool for Nationally Determined Contribution (NDC) planning.