<p>Accurate prediction of carbon emissions is fundamental to effective climate governance, yet it is inherently challenging due to the complex, multi-scale, and uncertain nature of emission data. To address these challenges, this study introduces a novel carbon emission prediction framework that integrates advanced computational techniques. The framework first employs a fuzzy entropy-constrained variational mode decomposition method for sophisticated signal denoising and pattern preservation. It then utilizes a temporally reinforced inductive graph attention network to capture intricate short and long term spatial–temporal dependencies. A pattern recognition system that combines clustering with adversarial training is incorporated to extract and leverage shared knowledge, while a forest copula architecture models nonlinear dependencies to generate robust probabilistic predictions. Experimental results demonstrate that this integrated approach achieves a 40.88% improvement in point prediction accuracy and a 45.82% enhancement in the reliability of interval predictions, significantly outperforming existing benchmark models. Furthermore, interpretability analyses validate the framework's capability in pinpointing the primary spatial–temporal drivers of carbon fluctuations and in disentangling the interactive relationships among various emission sources. This provides actionable insights for policymakers, establishing a new paradigm for reliable carbon emission prediction that successfully balances predictive accuracy, uncertainty quantification, and decision-making relevance.</p>