TyphoonWorld: A Residual Token World Model for Probabilistic Forecasting of Typhoon Imagery
Organizations: Nanyang Business School Nanyang Technological University, Singapore · National Academy of Governance Beijing, China · PBC School of Finance Tsinghua University, Beijing, China
Abstract
A satellite forecast of a typhoon should give a low-error picture of the next two days, together with an ensemble whose members are coherent fields and whose spread responds to the storm at hand. A small direct regression network already gives the lowest squared error on held-out storms, so a generative model has to earn its cost through the part of the future that the regression smooths away. We present TyphoonWorld, a discrete-token world model that generates only that part. Conditioned on tokens of the regression forecast, the latest frames and the best-track state at forecast time, it samples residual fields lead by lead, and members add these residuals to the regression center. We compare it under one protocol with per-pixel Gaussian networks, a diffusion model of the same residuals, residuals from other storms and an unconditional token world model, on 276 forecast windows from 46 held-out storms of Digital Typhoon. TyphoonWorld matches the squared error of the regression, lowers the CRPS of the area colder than 220 K, a direct measure of deep convection, by 14 to 39% against every other ensemble, and ties its spread to forecast error as closely as the Gaussian network, with a correlation of 0.78 across windows. It uses 56% of the parameters of the unconditional world model.
Figures & tables
| Point | Proper scores | Spread and error | |||||||
| Model | Params | MSE | CRPS | Energy | 220 K | 240 K | Core | Window | Map |
| Persistence | — | 0.0459 | 0.1504 | 2.683 | 0.0194 | 0.0395 | 0.0823 | — | — |
| ConvLSTM | 0.52M | 0.0368 | 0.1329 | 2.567 | 0.0482 | 0.1369 | 0.2015 | — | — |
| SimVP-style, closed loop | 2.18M | 0.0442 | 0.1556 | 2.868 | 0.0493 | 0.1288 | 0.2226 | — | — |
| PhyDNet | 2.93M | 0.0246 | 0.1185 | 2.012 | 0.0413 | 0.0972 | 0.0698 | — | — |
| TAU, chunked | 41.3M | 0.0259 | 0.1195 | 2.090 | 0.0420 | 0.0987 | 0.0750 | — | — |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Split | Years | Storms | Frames |
|---|---|---|---|
| Train | 1978–2018 | 998 | 170,760 |
| Validation | 2019–2020 | 51 | 9,783 |
| Test | 2021–2022 | 47 | 8,810 |
| Component | Setting | Value |
|---|---|---|
| Frame tokenizer | Codebook size and dimension | 1,024 and 256 |
| Spatial downsampling | to a grid | |
| EMA decay, commitment, SSIM weights | 0.99, 0.25, 0.3 | |
| Steps, batch size, learning rate | 20,000, 192, | |
| Residual tokenizer | Initialization, residual scale | frame tokenizer, 1.5 |
| Steps, windows and leads per batch | 14,000, 8 and 12 |
| Model | +1 h | +6 h | +12 h | +24 h | +36 h | +48 h |
|---|---|---|---|---|---|---|
| MSE of the point forecast | ||||||
| Persistence | 0.0081 | 0.0311 | 0.0406 | 0.0480 | 0.0551 | 0.0588 |
| ConvLSTM | 0.0038 | 0.0226 | 0.0364 | 0.0409 | 0.0421 | 0.0439 |
| SimVP-style, closed loop | 0.0032 | 0.0186 | 0.0346 | 0.0480 | 0.0594 | 0.0583 |
| PhyDNet | 0.0051 | 0.0166 | 0.0225 | 0.0260 | 0.0294 | 0.0298 |
| TAU, chunked | 0.0082 | 0.0186 | 0.0233 | 0.0270 | 0.0305 | 0.0320 |
| Ensemble | ||||||
|---|---|---|---|---|---|---|
| TyphoonWorld, , | 0.0920 | 0.0868 | 0.0851 | 0.0860 | 0.0884 | 0.0948 |
| TyphoonWorld, , | 0.0885 | 0.0835 | 0.0830 | 0.0853 | 0.0889 | 0.0973 |
| TyphoonWorld, , | 0.0872 | 0.0833 | 0.0839 | 0.0871 | 0.0914 | 0.1010 |
| TyphoonWorld, , | 0.0879 | 0.0833 | 0.0830 | 0.0854 | 0.0891 | 0.0977 |
| TyphoonWorld, , | 0.0891 | 0.0847 | 0.0842 | 0.0862 | 0.0895 | 0.0977 |
| TyphoonWorld, , | 0.0906 | 0.0855 | 0.0839 | 0.0848 | 0.0871 | 0.0938 |
| CRPS | Energy | 220 K | 240 K | Core | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Comparison | Won | Won | Won | Won | Won | |||||
| Gaussian U-Net | 0.20 | 0.001 | 1.00 | 0.001 | 0.91 | 0.001 | 0.96 | 0.001 | 1.00 | 0.001 |
| Gaussian U-Net, correlated | 0.17 | 0.001 | 0.59 | 0.771 | 0.72 | 0.001 | 0.87 | 0.001 | 0.50 | 0.696 |
| Direct U-Net + residuals | 1.00 | 0.001 | 0.48 | 0.778 | 0.93 | 0.001 | 0.91 | 0.001 | 0.39 | 0.251 |
| Diffusion residuals | 0.43 | 0.721 | 0.35 | 0.055 | 0.63 | 0.020 | 0.74 | 0.001 | 0.46 | 0.696 |
| Token rollouts | 0.74 | 0.001 | 0.96 | 0.001 | 0.89 | 0.001 | 0.96 | 0.001 | 0.85 | 0.001 |
| Point | Proper scores | Spread and error | ||||||
| Run | MSE | CRPS | Energy | 220 K | 240 K | Core | Window | Map |
| Seed 0 | 0.0242 | 0.0839 | 1.475 | 0.0115 | 0.0253 | 0.0501 | 0.78 | 0.25 |
| Seed 1 | 0.0242 | 0.0838 | 1.472 | 0.0113 | 0.0251 | 0.0499 | 0.80 | 0.25 |
| Seed 2 | 0.0242 | 0.0837 | 1.472 | 0.0114 | 0.0250 | 0.0496 | 0.80 | 0.25 |
| Mean of three seeds | 0.0242 | 0.0838 | 1.473 | 0.0114 | 0.0251 | 0.0499 | 0.79 | 0.25 |
| Standard deviation | 0.00001 | 0.00010 | 0.0014 | 0.00010 | 0.00016 | 0.00024 | 0.015 | 0.001 |
Probabilistic Forecasting
Probabilistic forecasting predicts distributions or events with quantified uncertainty to support reliable decisions under uncertainty.
See all 253 papers