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

Yuze Sun; Shiyi Wang; Jiancheng Pan; Die Wang; Andreas F. Prein; Wentao Luo; Linhan Jiang; Jie Wu

cs.LGSep 7, 2026

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

Authors: Yuze Sun, Shiyi Wang, Jiancheng Pan, Die Wang, Andreas F. Prein, Wentao Luo, Linhan Jiang, Jie Wu, +2 more

Organizations: 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 · School of Information and AI, Beijing Forestry University, Beijing, China · State Key Laboratory of Climate System Prediction and Risk Management/China Meteorological Administration Key Laboratory for Climate Prediction Studies, National Climate Centre, China Meteorological Administration, Beijing, 100081 · National Institute of Natural Hazards, Ministry of Emergency Management of China, Beijing 100085, China.

Abstract

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.

Figures & tables

Figure 1: Overview of the proposed PCSDiff framework. The framework consists of two major stages. Stage 1 is the PIMD - based dynamic multi - day bias correction module, which takes 10 - day lead - time ensemble precipitation forecasts as input, and employs the synoptic condition encoder, temporal history encoder, and precipitation intensity - aware multi - branch decoder (PIMD) to produce bias - corrected low - resolution precipitation. Stage 2 is the two - phase super - resolution module: Phase 1 applies a Conv - ResNet for regression - based preliminary upscaling, and Phase 2 further reconstructs fine - scale precipitation details via a conditional denoising diffusion model with denoising U - Net backbone, yielding the final high - resolution corrected precipitation outputs.

Figure 2: Performance metrics versus forecast lead time over China. Shown are raw ECMWF forecasts and two variants of our model using TP - only and full - atmospheric inputs. Subplots include continuous metrics (RMSE, MAE, ACC), event - based TP and CSI scores and spatial FSS score. Light - green shading denotes the performance gain of our model over raw ECMWF forecasts, evaluated on 1–10 - day forecasts over China.

Method	Lead time = 3 day					Lead time = 5 day
Method	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$
ECMWF	4.759	0.607	0.765	0.655	0.874	5.166	0.535	0.740	0.623	0.839
ViT ( Dosovitskiy et al. 2020 )	4.317	0.598	0.743	0.595	0.867	4.770	0.526	0.721	0.571	0.840
ConvLSTM ( Zhang and Di 2025 )	4.405	0.583	0.729	0.587	0.851	4.807	0.519	0.707	0.565	0.824
TransUNet ( Ye et al. 2026 )	4.344	0.593	0.731	0.579	0.851	4.782	0.524	0.711	0.558	0.828
CorrDiff ( Mardani et al. 2025 )	4.404	0.589	0.737	0.602	0.855	4.830	0.522	0.715	0.572	0.829

Table 1: Performance comparison against baseline methods at 3 - day and 5 - day forecast lead times over China region

Figure 3: Spatial precipitation distributions for three representative heavy - rainfall events. Each row corresponds to one case on different date and lead time. From left to right: raw ECMWF forecast, low - resolution bias - corrected output, high - resolution super - resolution result, and CMA - CRA ground truth.

Method	Lead time = 7 day					Lead time = 10 day
Method	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$
ECMWF	5.739	0.438	0.703	0.577	0.769	6.402	0.309	0.658	0.515	0.676
ViT ( Dosovitskiy et al. 2020 )	5.280	0.432	0.686	0.536	0.775	5.816	0.327	0.648	0.496	0.699
ConvLSTM ( Zhang and Di 2025 )	5.309	0.422	0.677	0.533	0.764	5.848	0.320	0.643	0.503	0.693
TransUnet ( Ye et al. 2026 )	5.292	0.427	0.679	0.526	0.765	5.828	0.323	0.643	0.488	0.694
CorrDiff ( Mardani et al. 2025 )	5.325	0.422	0.681	0.527	0.761	5.821	0.317	0.644	0.478	0.686

Table S1: Performance comparison against baseline methods at 7 - day and 10 - day forecast lead times over China region

Variable	Name	Pressure Levels	Role for total precipitation (tp)
2t	2m temperature	Single surface level	Determines precipitation phase
gh	geopotential height	850hPa, 700hPa, 500hPa	Characterizes circulation patterns
msl	mean sea - level pressure	Single surface level	Important for pressure systems and fronts
q	specific humidity	1000hPa, 925hPa, 850hPa, 700hPa, 500hPa	Represents atmospheric moisture content
sp	surface pressure	Single surface level	Important surface boundary condition
sst	sea surface temperature	Single surface level	Significant over ocean regions

Table S2: Input meteorological variables from ECMWF ensemble forecast.

Method	Lead time = 3 day					Lead time = 5 day
Method	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$
ECMWF	4.759	0.607	0.765	0.655	0.874	5.166	0.535	0.740	0.623	0.839
w/o other variables (only precipitation)	4.036	0.646	0.780	0.667	0.887	4.412	0.579	0.752	0.640	0.853
w/o TP loss (MSE Loss)	4.220	0.627	0.758	0.644	0.871	4.531	0.562	0.734	0.628	0.838
w/o TP loss (MSE + FSS Loss)	4.228	0.631	0.767	0.651	0.868	4.547	0.568	0.741	0.633	0.841
w/o PIMD module	4.190	0.638	0.769	0.653	0.872	4.516	0.574	0.743	0.636	0.840

Table S3: Ablation study of key components in PCSDiff over China region (Lead time = 3/5 day)

Method	Lead time = 7 day					Lead time = 10 day
Method	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$	RMSE $\downarrow$	ACC $\uparrow$	TP $\uparrow$	CSI $\uparrow$	FSS $\uparrow$
ECMWF	5.739	0.438	0.703	0.577	0.769	6.402	0.309	0.658	0.515	0.676
w/o other variables (only precipitation)	5.007	0.476	0.714	0.599	0.781	5.632	0.352	0.668	0.545	0.687
w/o TP loss (MSE Loss)	5.071	0.462	0.699	0.599	0.774	5.600	0.339	0.656	0.562	0.685
w/o TP loss (MSE + FSS Loss)	5.099	0.468	0.705	0.601	0.778	5.642	0.345	0.660	0.562	0.689
w/o PIMD module	5.074	0.472	0.707	0.596	0.775	5.613	0.348	0.662	0.559	0.683

Table S4: Ablation study of key components in PCSDiff over China region (Lead time = 7/10 day)

Category	Parameter	Value	Description
Training – Stage 1	Input channels	39	10 ensemble precipitation channels and 29 auxiliary atmospheric-variable channels
	Output channels	1
	PIMD branches	3	Number of parallel PIMD branches
	Base feature dimension	64	Initial feature dimension of the U-Net
	Dimension multipliers	(1, 2, 4, 8)	Feature multipliers at different network levels
	Batch size per GPU	1

Table S5: Training – Stage 1 Hyper - parameters.

Category	Parameter	Value	Description
Training – Stage 2 Phase 1	Input channels	1
	Output channels	1
	Base feature channels	64	Base feature width of the regression network
	Residual blocks	8
	Residual scaling	0.5	Residual-feature scaling factor
	Batch size	4

Table S6: Training – Stage 2 Hyper - parameters.

Category	Parameter	Value	Description
Dataset	Train set	2004–2021
	Test set	2022–2024
	Forecast data	ECMWF S2S	Ensemble forecasts and atmospheric variables
	Reference precipitation	CMA-CRA	High-resolution precipitation observations
	Stage-1 resolution	$1.5^{\circ}\rightarrow 1.5^{\circ}$	Forecast correction at the original resolution
	Stage-2 resolution	$1.5^{\circ}\rightarrow 0.25^{\circ}$	Downscaling from ECMWF forecasts to the CMA-CRA resolution

Table S7: Dataset and Hardware Configuration.

Observation ( $x_{i,j}$ )	Forecast ( $f_{i,j}$ )
	$\geq 1\,\text{mm/day}$	$<1\,\text{mm/day}$
$\geq 1\,\text{mm/day}$	NA (Hit)	NC (Miss)
$<1\,\text{mm/day}$	NB (False alarm)	ND (Correct negative)

Table S8: Contingency table for categorical precipitation verification (threshold:

1\,\text{mm/day}

)

Configuration	Ensemble Members	GPU Memory (GB)	End - to - end Latency (s)
Full sampling ( $N_{\text{step}}=50$ )	5	12.4	14.7
Full sampling ( $N_{\text{step}}=50$ )	10	18.7	38.2
Accelerated sampling ( $N_{\text{step}}=10$ )	5	11.8	3.6
Accelerated sampling ( $N_{\text{step}}=10$ )	10	17.2	13.1

Table S9: Inference performance of PCSDiff deployment pipeline under different settings. Latency denotes wall - clock time for complete 10 - day forecast inference.

Explore similar work

SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland
May 15, 2026Dan Assouline, Erwan Koch, Federico Amato +5Precipitation Forecasting Time Series Forecasting
Swiss Data Science Center, ETH, Zürich, Switzerland · Expertise Center for Climate Extremes, University of Lausanne, Lausanne, Switzerland · Faculty of Geosciences and Environment, University of Lausanne, Lausanne, Switzerland +1
Diffusion Precipitation Generation with Medium-range Weather Forecasting
Oct 9, 2026Guowen Li, Yang Liu, Mengxuan Chen +11Precipitation Forecasting Weather Forecasting
Tsinghua University · The Chinese University of Hong Kong · The Hong Kong University of S & T +5
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

May 15, 2026J/K move · Enter open · S save

SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland

Dan Assouline, Erwan Koch, Federico Amato, Filippo Quarenghi+4

Swiss Data Science Center, ETH, Zürich, Switzerland · Expertise Center for Climate Extremes, University of Lausanne, Lausanne, Switzerland · Faculty of Geosciences and Environment, University of Lausanne, Lausanne, Switzerland+1

Precipitation Forecasting Time Series Forecasting Probabilistic Forecasting Precipitation Downscaling Weather Forecasting

Open paper AskPDF arXiv

Skillful medium-range precipitation forecasting at kilometer scale remains challenging over complex terrain because precipitation arises from multiscale nonlinear processes that global models cannot explicitly resolve at affordable cost. Global AI weather models can produce skillful medium-range forecasts, but their native 0.25 degrees resolution limits direct use for local hazard applications. Statistical downscaling can help bridge this gap, yet existing approaches often struggle with state-dependent, and especially lead-time-dependent, biases in global forecasts. We introduce SwAIther-Precip, a lead-time-aware downscaling framework that converts coarse-resolution AIFS forecasts into probabilistic km-scale precipitation fields over Switzerland. First, a U-Net conditioned on lead time via feature-wise linear modulation deterministically corrects systematic biases at coarse resolution. This targeted correction enables a cheaper super-resolution stage conditioned only on corrected precipitation, allowing direct training on observations rather than on the full atmospheric state. A diffusion-based model then generates fine-scale spatial variability independently of lead time. Using AIFS forecasts and CombiPrecip radar-gauge observations, SwAIther-Precip reduces CRPS by 48% relative to raw AIFS. The generated fields reproduce observed spatial variability with spectral fidelity above 0.85 at large scales and 0.88 at small scales, corresponding to an effective resolution of approximately 4 km on a 1 km grid for lead times up to 5 days. Training across lead times further improves long-range performance, yielding a 13% CRPS reduction at 6 days relative to lead-time-specific models. These results show that explicitly correcting lead-time-dependent biases before generative super-resolution is key to efficient km-scale probabilistic downscaling of global AI precipitation forecasts.

Abstract

Explore similar work

SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Precipitation Downscaling Using Foundation Model-Conditioned Diffusion