Toward actionable insights for carbon neutrality: A sustainable spatio-temporal forecasting framework fusing socio-economic dependencies

Accurate carbon emission forecasting (CEF) is critical for achieving carbon neutrality. However, existing models often fail to capture complex socio-economic interdependencies and entail high computational carbon costs. To address these limitations, this study hypothesizes that integrating socio-economic theory with an adaptive and efficient neural architecture can simultaneously enhance prediction accuracy and computational sustainability. This study proposes a novel CEF framework that combines STIRPAT economic theory, a Graph Sparse Gating Decision Network (GSGDN), and a forest copula function. The framework first constructs multi-view socio-economic graphs via geographically weighted regression to represent regional dependencies. It then employs a GSGDN that fuses socio-economic and spatio-temporal information through dedicated modules while dynamically allocating computational resources via a sparse gating mechanism to reduce carbon footprint. Finally, a forest copula models cross-regional prediction-error dependencies to provide robust uncertainty quantification. Comprehensive evaluations show that the framework achieves a mean absolute percentage error (MAPE) of 6.619% with an operational carbon footprint of only 58.561 g/h, outperforming benchmark models in point and interval forecasting, as well as in high-emission and out-of-sample scenarios. These results demonstrate that the proposed framework offers a sustainable and precise decision-support tool for carbon governance, enabling differentiated regulation, proactive policy-making, and early-warning systems.