Forecasting¶
Forecast the temperature 24 hours ahead: at a weather station from its last 48 hours of measurements, or everywhere on Earth with a pre-trained global weather model. Dataset: Jena climate, ERA5 (WeatherBench 2).
Training¶
Make the model itself uncertainty-aware.
-
Bayesian weather forecasting: what the model knows and what it doesn't
Separate the noise in the weather from the model's own uncertainty, with Bayesian layers.
Conv1DFlipout·DenseFlipout·GaussianHead -
Probabilistic weather forecasting with a Gaussian head
Forecast tomorrow's temperature together with how sure the forecast is.
GaussianHead
Post-processing¶
Add calibrated uncertainty to an already trained model, without retraining it.
-
Conformal prediction intervals for a weather forecaster
Add intervals with a 90% coverage guarantee to any trained forecaster, after training.
SplitConformalRegressor -
Probabilistic weather forecasts from a pre-trained FourCastNet
Turn a pre-trained global weather model into 90% forecast intervals at every point on Earth, without any training.
SplitConformalRegressor