Preprint

ProtoCast

Interpretable Prototype-Conditioned
Adaptation for Weather Forecasting

Zhixuan Li1Hao Jiang1Chen Hui2Kai Geng2Weisi Lin1*

1Nanyang Technological University, Singapore2Nanjing University of Information Science and Technology, China

*Corresponding author

01 · Forecasts in focusGlobalForecast
Ground truthProtoCast
ProtoCast 12-hour temperature forecast from the main paperGround-truth temperature at the same forecast lead

2-m temperature · +12h · original 32 × 64 grid

01 / 03

A forecast.
A closer look.

Compare the same weather field, then look at the errors that separate two predictions.

Explore this example

GlobalForecast

18.30%

Lower temperature MSE ↓

0.9608 → 0.7850, compared with Met²Net.

TyphoonTracking

25.95%

Lower temperature MSE ↓

0.7467 → 0.5529, compared with TAU.

Local patterns.
Explicit contributions.

Group-specific prototype banks help adapt local features while accounting for cross-variable dependencies.

See one contribution

02 · Why prototypes?

What does one
prototype change?

Follow the paper’s T2M example from a matched pattern to a measurable contribution.

Drag to rotate · Pinch or +/− to zoom

01

Find a local match.

The current input is matched against learned prototypes. In this example, P27 and P31 provide the explicit pattern references.

P27 0.88P31 0.12

03 · Global forecasts

Same field.
See the difference.

Compare ground truth and a prediction in place. Drag the divider to inspect the same grid cells.

2-m temperature

+12h · original 32 × 64 grid

ProtoCast
Ground truthProtoCast
ProtoCast temperature forecastGround-truth temperature
201.61Field color scale shared across methods and cases321.17 K

Identical color scales are retained across methods and cases. UV10 shows wind speed; its error averages absolute U10 and V10 component errors. These selected examples are distinct from the aggregate benchmark results.

04 · Typhoon tracking

Keep the storm
in focus.

Replay the original Doksuri and Mawar forecast centers. Each map point is a +3h forecast from a separate six-hourly initialization.

10 initializations · ProtoCast vs. Met²Net

Swipe the figure to inspect the map and error curve ↔

Doksuri map and mean error curves from the paper
22 Jul 2023 · 12:00 UTC1 / 10
UTC+8

05 · How it works

Match locally.
Exchange across variables.

Historical features, temporal changes, and forecast context guide sparse prototype matching. Weighted low-rank operators produce explicit local contributions before gated cross-variable fusion.

EncodeMatchUpdateFuse

06 · Results

Every lead.
A closer comparison.

Inspect the exact four example-specific curves shown in the main paper. Each variable uses its own displayed test case.

UV10

Aggregate benchmarks

Global and regional forecasting.

MAE ↓ · lower is better · selected compared methods

ProtoCast achieves the lowest MAE on all four GlobalForecast variables.

The compact table displays MAE only. The full paper reports MSE, MAE, RMSE, and R² for all compared methods.

Citation

BibTeX

arXiv · 2026
@article{li2026protocast,
  title={Interpretable Prototype-Conditioned Adaptation for Weather Forecasting},
  author={Li, Zhixuan and Jiang, Hao and Hui, Chen and Geng, Kai and Lin, Weisi},
  journal={arXiv preprint},
  year={2026}
}
Back to top

Paper figure