AI and promote carbon reduction: exploring an optimal deep learning model to predict CO2 emissions under technological innovations and green finance

Abstract

Accurate prediction of carbon dioxide (CO₂) emissions is essential for effective carbon management and evidence-based climate policy. However, emission dynamics are inherently nonlinear and arise from complex interactions among technological, financial, energy, economic and demographic factors. This study develops an interpretable artificial-intelligence framework to predict CO₂ emissions across 15 Regional Comprehensive Economic Partnership (RCEP) economies during 2000–2024. Annual data were converted to quarterly frequency and decomposed using Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise. Fifteen neural-network and deep-learning architectures were evaluated using mean absolute error, mean squared error, root mean squared error, and training time. The Radial Basis Function network achieved the strongest overall performance, with MAE of 0.028, MSE of 0.002, RMSE of 0.040, and a training time of 00:01:12. Shapley Additive Explanations (SHAP) analysis and three-dimensional representation showed heterogeneous contributions from technological innovation, green finance for renewable energy, nonrenewable energy consumption, financial development, economic growth and population aging. Patent activity and R&D expenditure were predominantly associated with lower predicted CO₂ emissions, whereas green finance showed context-dependent predictive contributions. The findings demonstrate that a comparatively parsimonious RBF network can provide accurate and computationally efficient CO₂ prediction. By integrating signal decomposition, multi-model benchmarking and explainable AI, the proposed framework provides an interpretable decision-support tool for identifying emission patterns and informing targeted carbon-management and climate-mitigation strategies across RCEP economies.