TianQuan-S2S: A Subseasonal-to-Seasonal Global Weather Model via Incorporate Climatology State

TL;DR

TianQuan-S2S targets subseasonal-to-seasonal forecasting by integrating climatology into a patch-based embedding and enhancing variability capture with an uncertainty-augmented Transformer. The approach addresses model collapse and underutilization of climate states, delivering improved deterministic and ensemble predictions on ERA5 data for lead times $15$–$45$ days, outperforming ECMWF-S2S, FuXi-S2S, ClimaX, and climatology baselines. Ablation analyses confirm the value of explicit climatology incorporation and stochastic Transformer blocks, signaling a robust path toward more reliable long-range weather forecasts. The method holds potential for agriculture, energy, and emergency management, with future work aimed at higher resolution and additional environmental signals such as land, ocean, and sea ice data.

Abstract

Accurate Subseasonal-to-Seasonal (S2S) forecasting is vital for decision-making in agriculture, energy production, and emergency management. However, it remains a challenging and underexplored problem due to the chaotic nature of the weather system. Recent data-driven studies have shown promising results, but their performance is limited by the inadequate incorporation of climate states and a model tendency to degrade, progressively losing fine-scale details and yielding over-smoothed forecasts. To overcome these limitations, we propose TianQuan-S2S, a global S2S forecasting model that integrates initial weather states with climatological means via incorporating climatology into patch embedding and enhancing variability capture through an uncertainty-augmented Transformer. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset demonstrate that our model yields a significant improvement in both deterministic and ensemble forecasting over the climatology mean, traditional numerical methods, and data-driven models. Ablation studies empirically show the effectiveness of our model designs. Remarkably, our model outperforms skillful numerical ECMWF-S2S and advanced data-driven Fuxi-S2S in key meteorological variables.

Figures (9)

• Figure 1: Model collapse in Subseasonal-to-Seasonal (S2S) Forecasts. As the lead time increases, the forecast results for the same target time gradually exhibit the loss of data contours across the three regions, which can be considered a form of model collapse.
• Figure 2: The diagram of TianQuan-S2S framework. The input variables include initial state $\bm{X}$ and climatology $\bm{X}_{clim}$. After attention-based fusion, the features are enhanced and fused, then patchified and fed into the uncertainty-augmented model, where Gaussian noise is injected at each layer. Finally, predictions $\hat{\bm{X}}$ for days 15 to 45 are generated.
• Figure 3: TianQuan-S2S outperforms FuXi-S2S, ClimaX and ECMWF-S2S in deterministic forecasts on 1.40625°(a) and 5.625°(b) daily ERA5 datasets. The comparison involves four variables in terms of latitude-weighted RMSE (lower is better) and ACC (higher is better). The time range for the metric calculations is from 2017 to 2018.
• Figure 4: Visualization of forecast results on 1.40625° daily ERA5 data. The 30-day forecast of one upper-air variable (T850) and two surface variables (T2m and $\text{Wind}10$). For each case, the input time is 00:00 UTC on 15 February 2018.
• Figure 5: Scatter chart of prediction and the ground truth. The input time is 00:00 UTC on 15 February 2017 with the lead time of 25 days.