Forecasting HFC-R134a Trends Amidst Global Warming: A Comparative Study of Residual Transformers and Attention-Based RNNs for Middle-Term Projections

The continuous accumulation of hydrofluorocarbons (HFCs), particularly R134a (1,1,1,2-Tetrafluoroethane), poses a significant challenge to global climate mitigation efforts due to its high global warming potential. Accurate modeling of its atmospheric trajectory is complicated by nonstationary growth trends and hemispheric mixing asymmetries. This study introduces a novel deep learning framework to model and project R134a concentrations using high-resolution data sets from the Arctic (Barrow, Alaska) and Antarctic (South Pole) stations (1994–2025). Three parallel architectures are offered Attention-Based LSTM, Attention-Based GRU, and a Residual Transformer Encoder. To ensure robust generalization, hyperparameter tuning was conducted via Bayesian Optimization using an expanding window cross-validation strategy. Empirical results demonstrate that the Residual Transformer significantly outperforms recurrent architectures, achieving the lowest mean absolute error (MAE < 0.34 pmol/mol) and highest determination coefficient of 0.995 across both poles. A crucial finding is the distinct hemispheric disparity; models applied to the Antarctic data set exhibited superior stability and lower stochastic volatility compared to the Arctic, reflecting the impact of anthropogenic source proximity on Northern hemisphere atmospheric dynamics. Furthermore, a computational complexity analysis reveals that while the attention-LSTM is the most resource-efficient suitable for edge deployment, the residual transformer offers the optimal accuracy-to-cost ratio, surpassing the computationally expensive Attention-GRU. Finally, recursive long-term projections (2025–2035) indicate a continued monotonic increase in global R134a concentrations, serving as a baseline reference scenario to support future environmental assessments and broader atmospheric monitoring efforts.