Explainable and Uncertainty‐Aware Neural Modeling of Industrial CO 2 Emissions With Economic, Energy, and Institutional Interaction Dynamics

ABSTRACT This study develops an explainable and uncertainty‐aware neural framework for forecasting industrial carbon dioxide (CO 2 ) emissions in a heterogeneous global panel. The analytical panel integrates EDGAR, WDI, and V‐Dem data for 5138 country–year observations from 165 countries during 1990–2024. The proposed dual‐input neural network combines standardized economic, energy, and institutional interaction and lagged features with a learned country embedding to represent persistent cross‐national heterogeneity. The framework is formulated as conditional (covariate‐assisted) forecasting. The target enters only through its 1‐year lag, whereas economic, energy, and institutional covariates are taken for the forecast year. The target is log‐transformed industrial CO 2 emissions, and evaluation uses an expanding‐window rolling‐origin design in addition to time‐blocked fivefold cross‐validation, a strictly temporal hold‐out, benchmark comparison, ablation analysis, and uncertainty calibration. The full model achieved a mean cross‐validation R 2 of 0.9649 (SD = 0.0034), root mean squared error (RMSE) of 0.4069 (SD = 0.0176), and MAE of 0.3272 (SD = 0.0275). In rolling‐origin evaluation for 2010–2024, it retained a mean R 2 of 0.9446 and a mean RMSE of 0.4798. It also outperformed other benchmarks, including XGBoost, random forest, long short‐term memory, linear regression, ridge, persistence, and autoregressive distributed lag, on the reported temporal test set. The ablation results indicate that governance features and country embeddings enhance the model's predictive ability, while the institutional interaction terms alone do not necessarily reduce RMSE. Overall, the model achieves a 53.80% reduction in RMSE compared with the baseline specification. Lagged industrial emissions are always the most important predictive factor for all four methods of SHAP, LIME, PFI, and IG, and the other factors, in order of importance, are income, renewable energy, and governance‐conditioned interactions. Initial interval diagnostics indicated that 0.801 empirical coverage was achieved for nominal 0.90 intervals. The chronological split‐conformal recalibration resulted in 0.873 empirical coverage, which remains below the nominal level, and the intervals should be used as preliminary before high‐stakes use, such as policy deployment. The model was trained and tested only on the countries of interest. The framework offers a clear and transparent predictive tool for analyzing industrial‐emissions pathways and makes a clear distinction between predictive association and causal effect.