DS-HBI: Dual-stream fusion forecasting model with historical backfilling imputation

Abstract

Deep learning models demonstrate significant potential for atmospheric carbon concentration forecasting, yet confront dual challenges of pervasive data missingness in real-world monitoring scenarios and intricate multivariate dynamic interactions. This paper proposes a Dual-Stream fusion forecasting model with Historical Backfilling Imputation (DS-HBI), a parallel architectural framework that resolves these challenges through dual-modal complementary pathways. The first pathway processed raw incomplete sequences via masked self-attention to capture intrinsic patterns without imputation bias. In contrast, the second integrates dynamic time warping (DTW) and probabilistic imputation to reconstruct temporally consistent data. A gated attention mechanism dynamically fuses both streams, adaptively balancing their contributions to jointly capture multi-scale temporal features, including long-term trends and abrupt changes, while ensuring robustness under severe data missingness. Evaluated on multi-site Total Carbon Column Observing Network (TCCON) data, DS-HBI demonstrates superior performance in predicting ${\text{XCO}}_2$ and ${\text{XCH}}_4$, significantly reducing prediction errors compared to baseline methods. The model particularly excels in high missing-rate scenarios, with ablation studies confirming the necessity of its dual-stream design and hybrid imputation strategy.