Carbon emission causal discovery and multi-step forecasting using spatiotemporal information

Spurred by the urgency to combat climate change, lowering carbon emissions has become a critical global concern. Nevertheless, interregional emission causal discovery and emission forecasting, potent decision tools for reducing emissions, have been insufficiently researched. This study aims to bridge these gaps by unveiling two innovative algorithms. The first, a causal discovery algorithm, adeptly navigates the complexities of emission interactions among regions, utilizing matrix decomposition and diffusion models for nuanced insights. The second, a spatiotemporal forecasting algorithm, leverages hyperbolic graph neural networks and ordinary differential neural networks to refine predictions of regional carbon emissions. The efficacy of the proposed algorithms is verified via meticulous verification, including Taylor statistics, predictive errors, Diebold Mariano tests, and ablation experiments. This study identifies key features of the carbon emission network, including 32 regions in the Chinese mainland, and offers insights into joint reduction policies grounded in algorithm effectiveness. This research catalyzes theoretical progress in carbon emission causality analysis and forecasting within the context of the ongoing climate crisis while also facilitating the decision-making science of artificial intelligence. Furthermore, it offers practical applications in environmental protection strategies, ecological assessments, and advancing global carbon neutrality via information science.