Mlulti-Scale Temporal Integration for Enhanced Greenhouse Gas Forecasting: Advancing Climate Sustainability

Accurate forecasting of greenhouse gas (GHG) concentrations is crucial for climate policy evaluation, mitigation strategy formulation, and global sustainable development. Precise predictions of GHG concentrations provide early warnings of critical climate thresholds, enabling proactive policy adjustments, sustainable industrial transitions, and timely societal actions to ensure long-term ecological security and the well-being of future generations. Current modeling approaches face challenges in capturing multi-scale temporal patterns in GHG data, with traditional methods like Global Circulation Models (GCMs) and conventional machine learning models limited in simulating long-term variations and multi-scale patterns. To address these limitations, this study proposes MGGTSP, a novel multi-encoder framework that integrates daily and monthly data using an Input Attention encoder, an Autoformer encoder, and a Temporal Attention mechanism. Evaluated on NOAA datasets from Mauna Loa, Barrow, American Samoa, and Antarctica over five decades, MGGTSP outperforms 14 baseline models, achieving an R² of 0.9627 and a MAPE of 1.47%, demonstrating exceptional predictive accuracy. In multi-step forecasting, it shows only a 3.3% reduction in R² over ten steps, significantly outperforming Transformer-based models in separating short-term fluctuations from long-term trends. This research provides a robust tool for cross-scale climate prediction and policy formulation, enabling precise climate policy targeting, low-carbon transitions, sustainable resource use, and enhanced ecosystem resilience. It also advances the understanding of multi-scale data processing in climate science, supporting informed decision-making for sustainable development initiatives.