Skillful High-Resolution Ensemble Precipitation Forecasting with an Integrated Deep Learning Framework

TL;DR

This work tackles the challenge of producing skillful, high-resolution precipitation forecasts by introducing a physics-inspired hybrid DL framework that couples a deterministic meso-scale predictor with a probabilistic latent-diffusion component to represent sub-grid uncertainties. The deterministic branch leverages a $3$D Swin-Transformer to capture mesoscale evolution, while a conditional latent diffusion model, aided by a VAE, generates residual convective-scale precipitation and yields ensemble forecasts. Trained on ERA5 and CMPA data and validated with real-time ECMWF forecasts, the method achieves nearly unbiased ensemble distributions, improved representation of extreme events, and robust real-time CSI performance up to 5 days. The approach demonstrates practical potential for operational high-resolution forecasting by delivering sharp, calibrated ensembles and better alignment with observations than traditional reanalysis benchmarks.

Abstract

High-resolution precipitation forecasts are crucial for providing accurate weather prediction and supporting effective responses to extreme weather events. Traditional numerical models struggle with stochastic subgrid-scale processes, while recent deep learning models often produce blurry results. To address these challenges, we propose a physics-inspired deep learning framework for high-resolution (0.05\textdegree{} $\times$ 0.05\textdegree{}) ensemble precipitation forecasting. Trained on ERA5 and CMPA high-resolution precipitation datasets, the framework integrates deterministic and probabilistic components. The deterministic model, based on a 3D SwinTransformer, captures average precipitation at mesoscale resolution and incorporates strategies to enhance performance, particularly for moderate to heavy rainfall. The probabilistic model employs conditional diffusion in latent space to account for uncertainties in residual precipitation at convective scales. During inference, ensemble members are generated by repeatedly sampling latent variables, enabling the model to represent precipitation uncertainty. Our model significantly enhances spatial resolution and forecast accuracy. Rank histogram shows that the ensemble system is reliable and unbiased. In a case study of heavy precipitation in southern China, the model outputs align more closely with observed precipitation distributions than ERA5, demonstrating superior capability in capturing extreme precipitation events. Additionally, 5-day real-time forecasts show good performance in terms of CSI scores.

Figures (9)

• Figure 1: Decomposition of high-resolution precipitation
• Figure 2: Overall model architeture. $X_{t-1}/X_t$ includes surface variables (T2m, U10m, V10m, MSLP) and upper-air variables at 13 pressure levels (T, U, V, SP, Z) at time $t-1/t$. For the deterministic model, the inputs are $X_{t-1}$ and $X_t$, while for the probabilistic model, the input is $X_t$. The model is trained on ERA5 data but can take forecasts from other models during inference. $\overline{TP}$ is the mean precipitation at time t, $TP'$ is the residual precipitation. $Z$ is the latent space, and $Z_T$ is the latent variable at diffusion step T. The circle with “C” represents concatenation, black arrows represent the training process, and gray dashed lines indicate the inference process.
• Figure 3: Conditional diffusion model architecture peeblesScalableDiffusionModels2023. We introduce conditioning inputs by using the atmospheric state and mean precipitation.
• Figure 4: CSI of ablation studies against multiple precipitation thresholds, the experiment configuration referred Table \ref{['tab:ablation_study']}
• Figure 5: A case on 2021-08-01 01:00. The first row shows the ground truth, while the second row shows the predictions. The leftmost image in the second row represents the mean precipitation generated using ERA5 as input, the middle image (Residual_Member1) represents one member of the residual precipitation, and the rightmost image (Ensemble_PM) shows the ensemble mean precipitation obtained by probability matching ebert2001 across all ensemble members.