To address the challenges of industrial carbon emissions forecasting (multicollinearity and complex interdependencies), we propose a graph-based deep learning framework, DGL. Using EDGAR v8.0 data, we demonstrate excellent predictive performance across countries and industry sectors, achieving over 15% error reduction. We use GNN to model inter-industry relationships and temporal transformers to learn long-term patterns. Attention weighting and causal analysis are used to maintain interpretability and demonstrate policy relevance.
