High-Resolution Probabilistic Data-Driven Weather Modeling with a Stretched-Grid

TL;DR

This work introduces Bris CRPS-FFT, a high-resolution probabilistic data-driven weather model that uses a stretched-grid to provide 2.5 km regional detail within a global 31 km framework. It deploys a graph neural network with stochastic latent noise and a loss function combining global point-wise CRPS and a regional spectral CRPS term to enforce spatial coherence across scales. The model delivers competitive forecasts compared to MEPS and improves field coherence over MSE- and CRPS-only approaches, with spectral loss helping to align energy spectra at large scales. The approach enables fast, high-resolution probabilistic forecasts suitable for operational use and extreme-weather guidance, with future work targeting hourly forecasts and further spectral-energy refinements.

Abstract

We present a probabilistic data-driven weather model capable of providing an ensemble of high spatial resolution realizations of 87 variables at arbitrary forecast length and ensemble size. The model uses a stretched grid, dedicating 2.5 km resolution to a region of interest, and 31 km resolution elsewhere. Based on a stochastic encoder-decoder architecture, the model is trained using a loss function based on the Continuous Ranked Probability Score (CRPS) evaluated point-wise in real and spectral space. The spectral loss components is shown to be necessary to create fields that are spatially coherent. The model is compared to high-resolution operational numerical weather prediction forecasts from the MetCoOp Ensemble Prediction System (MEPS), showing competitive forecasts when evaluated against observations from surface weather stations. The model produced fields that are more spatially coherent than mean squared error based models and CRPS based models without the spectral component in the loss.

Figures (8)

• Figure 1: Schematic diagram of the ensemble training. (a) The preceding ($t_{-6\mathrm{h}}$) and the current ($t_{0\mathrm{h}}$) states from MEPS and ERA5 are used as inputs to the model. The inputs are replicated $M$ times, where $M$ is the number of ensemble members during training. (b) The model has an encoder-processor-decoder architecture, where noise ($z$) is injected into the latent space for stochasticity. (c) The predictions are evaluated against targets using a point-wise and spectral training objective.
• Figure 2: Precipitation fields over parts of the Nordics at 2.5 km horizontal resolution. Each panel represent different forecasting models: (a) The control member of the NWP modelling system meps. (b) Bris MSE, as described in nipen2025. (c) Bris CRPS, trained with point-wise crps only. (d) Bris CRPS-FFT trained with point-wise and spectral crps. Forecasts are initialized at 2022-06-01T00Z and have a lead time of 18 hours.
• Figure 3: Same as Fig. \ref{['fig:fields']}, but zoomed in over southern Norway.
• Figure 4: Quantile-quantile (QQ) plot aggregated over lead times 24h--60h. Forecasts are plotted as a function of observations, with the ideal case (gray line) being when the forecasts equal the observations. (a) is precipitation and (b) is wind speed.
• Figure 5: Discrete cosine transform (DCT) of 10m wind speed as a function of wave length for lead times +6h, and +60h (upper row), and the corresponding logarithmic difference with respect to meps (lower row). The x-axis is flipped such that the wave lengths increase to the left. The power spectra are averaged from 2022-06-01T00Z to 2022-07-01T00Z.