Diffusion Precipitation Generation with Medium-range Weather Forecasting

Guowen Li; Yang Liu; Mengxuan Chen; Jiashun Cheng; Haoyuan Liang; Shilei Cao; Yujie Wang; Juepeng Zheng

cs.LGOct 9, 2026

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Authors: Guowen Li, Yang Liu, Mengxuan Chen, Jiashun Cheng, Haoyuan Liang, Shilei Cao, Yujie Wang, Juepeng Zheng, +6 more

Organizations: Tsinghua University · The Chinese University of Hong Kong · The Hong Kong University of S & T · Sun Yat-Sen University · NSCC-SZ · China Meteorological Administration · GBA Weather Research Center · Huawei Technologies Co., Ltd

Abstract

Accurate daily precipitation forecasts are important for weather-sensitive decisions. Although recent Weather Foundation Models (WFMs) have achieved significant progress, they are trained on reanalysis data, which exhibit regional and intensity-dependent biases. Since retraining WFMs on observation-based targets is costly, we study Dynamics-Precipitation Decoupling, a plug-and-play framework that uses a separately trained generator to map non-precipitation forecasts from a frozen weather model to observation-aligned daily precipitation. Its key challenges are to capture heterogeneous rainfall distributions and align daily accumulations with atmospheric signals that evolve across space and time. To address them, we train a diffusion generator conditional on atmospheric states and satellite-gauge estimates of precipitation. A multi-scale recurrent encoder summarizes atmospheric evolution, while cross-attention aligns atmospheric and precipitation features during denoising. Extensive evaluation of various backbones, including Pangu-Weather, GraphCast, and HRES, on global and regional forecasting shows improved daily precipitation accuracy and event detection, with better representation of localized accumulation patterns. Global ensemble evaluation shows lower Brier scores than ENS and training-from-scratch GenCast at all reported lead times, with lower CRPS in most comparisons. Remarkably, a heavy-rainfall case study using 2,288 stations in China further supports improved daily accumulation patterns and event detection against ground observations.

Figures & tables

Figure 1: Motivation and overview. (a) Consecutive daily precipitation fields exhibit substantial variability. (b) Directly plugging the generator into forecast models outperforms forecast-field fine-tuning. (c) The encoder compresses atmospheric variables into spatiotemporal features, followed by cross-field alignment with precipitation to guide generation. Together, these observations motivate our Dynamics–Precipitation Decoupling paradigm and direct plug-in design for precipitation generation.

Figure 2: Overview of our two-step framework. The workflow cycles as follows: (1) An AI/NWP model produces two 12h autoregressive forecasts $(\bm{\hat{X}}_{t-12},\bm{\hat{X}}_{t})$ , which the Encoder converts into the features. (2) Guided by these features, the Denoiser generates a 24h accumulated precipitation forecast $\bm{\hat{Y}}_{t}$ via a $K$ -step reverse diffusion process. The details are shown in the bottom right.

Figure 3: Global

1.5^{\circ}

deterministic skill over the 2022 test set for forecast backbones, their variants with our framework, and external system references. IMERG is used as the verification reference.

Metrics	Models	Lead Time (Days)
Metrics	Models	1	2	3	4	5	6	7
CRPS $\downarrow$	ENS	1.528	1.740	1.824	1.975	2.161	2.346	2.484
	GenCast	1.512	1.651	1.778	1.943	2.147	2.255	2.368
	Pangu+ Ours	1.483	1.619	1.756	1.910	2.061	2.208	2.336
	GraphCast+ Ours	1.477	1.569	1.660	1.769	1.898	2.035	2.180
	HRES+ Ours	1.525	1.637	1.747	1.867	2.006	2.156	2.302

Table 1: Probabilistic forecasting for a 50-member ensemble.

Figure 4: Spatial comparison of Day-3 precipitation forecasts from the baseline models, their variants with our framework, and the IMERG reference. Boxes A–C highlight regional differences.

Figure 5: Regional evaluation over the United States, Southern Africa, and Australia. Left: RMSE and CSI across lead times. Right: precipitation fields from WFMs, their variants, and Observations.

Figure 6: Station-based evaluation over North China in July 2022, initialized on June 29.

Figure 7: (a) Impact of sampling steps. RMSE and ACC approach saturation at 100 steps. (b) Relative performance changes under Gaussian perturbed inputs; red indicates degradation.

Condition	Temp	Align	Generator	ACC $\uparrow$	RMSE $\downarrow$	HSS $\uparrow$	CSI ${}_{0.1}\uparrow$	CSI ${}_{1}\uparrow$	CSI ${}_{4}\uparrow$
GraphCast	–	–	Native	0.6091	6.120	0.5347	0.5097	0.4928	0.4271
	GRU	–	Diff.	0.6217	5.924	0.5914	0.5612	0.5267	0.4826
	LSTM	–	Diff.	0.5942	6.318	0.5165	0.4942	0.4725	0.3945
	–	Attn.	Diff.	0.6571	5.721	0.6213	0.6158	0.5572	0.5185
	–	Gated	Diff.	0.6126	6.067	0.5678	0.5367	0.5126	0.4438
	–	–	U-Net	0.4071	9.840	0.6268	0.6458	0.5464	0.4535

Table 2: Day-1 configuration comparison. “–” denotes an omitted proposed module; U-Net and CorrDiff replace the complete generator.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Type	Variable	Abbreviation	Levels
Static	Land–sea mask	LSM	–
Static	Orography	ORG	–
Surface	2-m temperature	T2m	–
Surface	10-m zonal wind	U10	–
Surface	10-m meridional wind	V10	–
Surface	Mean sea-level pressure	MSLP	–

Appendix

Table 3: ERA5 variable–level interface used in the experiments.

Figure 8: Computational overhead of our framework on Pangu-Weather and GraphCast in terms of parameters, inference time, and peak GPU memory. Pangu-Weather parameter counts include all four forecast models.

Figure 9: Cross-backbone reuse and fine-tuning diagnostic.

Figure 10: Temporal-density ablation under GraphCast conditioning: forecast accuracy and computational overhead with two-state and four-state atmospheric inputs.

Figure 11: Global precipitation fields at lead times of 1, 3, 5, and 7 days, with IMERG as the reference. Forecasts are initialized at 00:00 UTC on July 6, 2022.

Figure 12: Station-based comparison across three valid times, with forecasts initialized on June 29, 2022.

Explore similar work

PCSDiff: Diffusion-Based Bias Correction and Super Resolution Toward Practical Operational Medium-Term Precipitation Forecast
Sep 7, 2026Yuze Sun, Shiyi Wang, Jiancheng Pan +7Precipitation Forecasting Precipitation Downscaling
Department of Earth System Science, Ministry of Education Key Laboratory for Earth System Modelling, Institute for Global Change Studies, Tsinghua University, Beijing, China · Huawei Technologies Co., Ltd · Institute for Atmospheric and Climate Science, ETH Zürich, 8092 Zürich, Switzerland +3
Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion
Oct 8, 2026Jannik Wiese, Johannes Schusterbauer, Tommaso Martorella +1Precipitation Nowcasting Probabilistic Forecasting
Precipitation Downscaling Using Foundation Model-Conditioned Diffusion
Aug 26, 2026Victor Nascimento Ribeiro, Jorge Guevara, Jorge Sebastian Moraga +7Precipitation Downscaling Conditional Diffusion Models
IBM Research · Fathom · STFC Hartree Centre

Sep 7, 2026J/K move · Enter open · S save

PCSDiff: Diffusion-Based Bias Correction and Super Resolution Toward Practical Operational Medium-Term Precipitation Forecast

Yuze Sun, Shiyi Wang, Jiancheng Pan, Die Wang+6

Department of Earth System Science, Ministry of Education Key Laboratory for Earth System Modelling, Institute for Global Change Studies, Tsinghua University, Beijing, China · Huawei Technologies Co., Ltd · Institute for Atmospheric and Climate Science, ETH Zürich, 8092 Zürich, Switzerland+3

Precipitation Forecasting Precipitation Downscaling Spatiotemporal Forecasting Conditional Diffusion Models

Open paper AskPDF arXiv

Medium-range precipitation forecasts are impaired by persistent systematic biases, lead-time-dependent error accumulation, and coarse spatial resolution, restricting their reliability for flood-drought risk assessment. Existing AI correction techniques lack dedicated modeling for multi-day dynamic bias evolution and proper meteorological constraints, often generating over-smoothed rainfall structures, and cannot meet operational deployment demands. This work introduces PCSDiff, a cascaded task-decoupled diffusion framework targeting 10-day precipitation bias correction and downscaling. To jointly counteract temporal error drifts and reconstruct physically plausible local precipitation details, PCSDiff integrates the Precipitation Intensity-aware Multi-branch Decoder (PIMD) module for dynamic multi-day error mitigation using synoptic-temporal features, followed by a two-phase conditional diffusion super-resolution module to restore fine-scale precipitation patterns. Evaluated against CMA-CRA observations over China after global-data training, PCSDiff cuts RMSE by 16.1% and lifts ACC by 13.9% relative to raw ECMWF forecasts at 3-10-day lead times, and consistently outperforms mainstream deep-learning baselines on both general and extreme-precipitation metrics. Benefiting from a streaming inference pipeline, our method achieves low-latency rolling forecasting for practical meteorological operations.

Abstract

Explore similar work

PCSDiff: Diffusion-Based Bias Correction and Super Resolution Toward Practical Operational Medium-Term Precipitation Forecast

Just Weather Scoring: Efficient End-to-end Nowcasting with Distributional Diffusion

Precipitation Downscaling Using Foundation Model-Conditioned Diffusion