PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment

Guowen Li, Yang Liu, Yujie Wang, Qiuyan Sun, Haoyuan Liang, Juepeng Zheng, Hong Cheng, Haohuan Fu

arXiv:2610.12401v1cs.LG

TL;DR

Regional forecasting needs local detail, while existing data-driven approaches either rely on global numerical forecasts or require training global components. ScaleCast aligns global guidance with regional locations and interactions; its forecasts improve across variables and drivers, and it outperforms CERRA NWP on all nine variables with GraphCast.

  • Problem

    High-resolution regional forecasting needs large-scale guidance, but pretrained global forecasts have coarse grids and existing approaches may depend on upstream numerical forecasts or train additional global components.

    Source: §1 · pp. 1–2

  • Method

    ScaleCast maps joint global–regional representations to regional positions and uses attention and gated updates to evolve the regional state under global forecast guidance.

    Source: §1 · p. 2, §3.2 · p. 5, §5 · p. 9

  • Results

    ScaleCast improves each global driver’s forecasts across available variables; with GraphCast it beats CERRA NWP on all nine, reducing MSLP RMSE by 8.9 Pa.

    Source: §4.2 · p. 7

  • Takeaways & Limitations

    One trained regional model can use Pangu-Weather, GraphCast, or HRES without driver-specific retraining, with improved forecasts across surface and upper-air variables.

    Source: §1 · p. 3, §5 · p. 9

    The study evaluates forecasts over Europe and HRRR transfer, and its results do not constitute operational certification or replace established warning procedures.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Kilometer-scale regional weather forecasting is essential for local weather warnings and weather-sensitive decisions. Existing data-driven approaches often rely on numerical forecasts for large-scale guidance or require additional training of global forecasting components. Pretrained global weather models offer an efficient source of large-scale forecasts, motivating their reuse to guide high-resolution regional prediction. However, this coupling requires aligning global and regional representations across different grids and integrating global guidance with local interactions to advance regional states. We propose ScaleCast, a regional forecasting framework that addresses these challenges through Global-Regional Alignment. Its Global-Regional Conversion module aligns joint global and regional representations with regional locations, while the Global-Regional Alignment and Dynamics block combines aligned guidance with regional neighborhood interactions. Experiments using ERA5 global analyses on a 0.25-degree grid and CERRA regional reanalysis at 5.5 km spacing demonstrate improved regional forecasts across surface and upper-air variables, with a single trained model supporting multiple global forecast drivers (i.e., Pangu-Weather, GraphCast, and HRES) without specific retraining. Fine-tuning on HRRR at 3 km spacing further demonstrates the framework's adaptability to a different regional domain and spatial resolution. Windstorm case studies show improved cyclone positioning and core-pressure estimates, while comparisons with HadISD station observations show closer agreement with local temperature and humidity changes.

1 INTRODUCTION

ScaleCast reuses pretrained global forecasts as large-scale guidance for high-resolution regional prediction, addressing the mismatch between coarse global and fine regional grids. Experiments report improved forecasts across variables and evaluation settings.

1 INTRODUCTION

  • Regional forecasts support local warnings and decisions, but conventional limited-area NWP requires computationally demanding integration and global forecasts for boundary conditions.These dependencies create trade-offs among resolution, update frequency, and delivery time.Source: §1 · p. 1
  • Prior data-driven approaches either use external fields for regional prediction or develop global forecasting components within the pipeline, with the latter incurring additional training costs.Pretrained global models offer an efficient source of large-scale forecasts, but their coarse grids do not resolve kilometer-scale regional detail.Source: §1 · pp. 1–2
  • GRConv maps joint global–regional representations to regional positions, while GRAD combines aligned guidance with regional attention and updates to advance the regional state.The framework is trained on ERA5–CERRA data and supports Pangu-Weather, GraphCast, and HRES without driver-specific retraining.Source: §1 · p. 2
  • ScaleCast improves surface and upper-air forecasts through +30 h, outperforms CERRA NWP on all nine variables with GraphCast, and shows closer agreement with station observations and lower windstorm errors.The windstorm case reports lower cyclone-position errors and core-pressure bias.Source: §1 · pp. 2–3

2 RELATED WORK

Related work spans learned global forecasting, several approaches to regional coupling, and post-processing or downscaling. ScaleCast distinguishes itself by learning cross-grid feature coupling to predict an increment of an existing regional state.

Global weather forecasting

  • Pretrained global forecasters, including Pangu-Weather and GraphCast, motivate using their forecast fields as regional guidance rather than retraining a global forecaster.Earlier work includes graph and Fourier operators, while later systems use transformers, multiscale graphs, hybrid modeling, pretraining, and diffusion.Source: §2 · p. 3

Regional weather forecasting

  • Regional methods differ in where external information enters and how global and regional scales interact; ScaleCast projects concatenated tokens onto regional positions before modeling neighborhood interactions.The framework diagram depicts its training pipeline, GRAD blocks, cross-grid token alignment, and windowed attention.Source: §2 · p. 3, Figure 2 · p. 4

Regional post-processing and downscaling

  • Unlike post-processing or coarse-to-fine reconstruction, ScaleCast predicts an increment of the existing regional state through learned cross-grid feature coupling.At inference, forecasts from pretrained global weather models provide boundary guidance for regional forecasting.Source: §2 · p. 3

3 METHODOLOGY

ScaleCast learns autoregressive regional forecasts conditioned on global forecasts, with terrain and the current regional state as inputs. GRConv aligns global and regional features, while GRAD updates regional features using neighborhood interactions and gated residual branches.

3 METHODOLOGY

  • ScaleCast learns regional forecasts autoregressively from a frozen global model’s trajectory, using 3-hour steps through 72 hours.Source: §3 · p. 4
  • Training uses global analysis, the current regional state, and terrain for single-step, per-variable MSE supervision, with FAMO adjusting variable weights by loss progress.Source: §3 · p. 4

3.1 GLOBAL–REGIONAL TOKEN ALIGNMENT

  • Because the global and regional grids lack one-to-one patch correspondence, GRConv projects joint global–regional features onto regional token positions while retaining a direct regional feature path.Source: §3.1 · pp. 4–5

3.2 REGIONAL INTERACTION AND STATE UPDATE

  • GRAD applies input-dependent windowed attention to aligned features and surrounding regional context, then uses channel-wise gates on attention and MLP residual branches while preserving the identity path.Source: §3.2 · p. 5
  • The decoder predicts changes relative to the current regional forecast, then uses a shared upsampler and variable-specific heads to produce regional-grid predictions.Source: §3.2 · p. 5

4 EXPERIMENTS

Experiments evaluate ScaleCast on ERA5–CERRA data against global and regional forecasting baselines, then examine module choices and case studies. Results show gains across variables and global drivers, with additional evaluations of storm forecasts, station observations, and transfer to HRRR.

4.1 EXPERIMENTAL SETUP

  • The experiments pair ERA5 atmospheric states with CERRA targets on a 512 × 512 grid at approximately 5.5 km spacing, and compare global, regional, and numerical forecasting baselines.Evaluation uses nine regional variables and includes Pangu-Weather, GraphCast, HRES, and regional forecasting systems.Source: §4.1 · p. 6

4.2 OVERALL PERFORMANCE

  • With GraphCast, ScaleCast outperforms all non-ScaleCast baselines on nine variables, and one shared checkpoint improves all three global drivers without driver-specific retraining.Across the three drivers, T2m RMSE falls by 0.503, 0.505, and 0.389 K for Pangu-Weather, GraphCast, and HRES, respectively.Source: §4.2 · p. 7

4.3 MODULE COMPARISONS AND ABLATIONS

  • GRConv outperforms three alternative spatial-alignment approaches on all nine variables, while decoder comparisons favor residual prediction over direct prediction across all nine variables.Attention variants have variable-specific advantages; low-rank residual heads improve six of nine variables over shared decoding.Source: §4.3 · p. 7, Table 2 · p. 8

4.4 CASE STUDY

  • In event 1804, GraphCast-driven ScaleCast has the lowest displayed position errors from +36 to +60 h and a 53.6 Pa core-pressure bias, while station comparisons show closer agreement on selected temperature and humidity changes.At S2, it follows the observed temperature decline more closely around +18 to +30 h and better captures high humidity around +30 to +42 h, but remains too humid during sharp drying around +57 to +69 h. On HRRR, SC+GraphCast reduces +3 h T2m and MSLMA RMSE by 4.3% and 4.2% versus StormCast, and by 1.5% and 5.1% versus HRRR NWP; this is a cross-dataset transfer test, not information-matched operational skill.Source: §4.4 · pp. 8–9

5 CONCLUSION

ScaleCast uses global–regional alignment to guide kilometer-scale regional forecasts. Its evaluations show improvements across variables and drivers, with further evidence from storm, station, and HRRR transfer experiments.

5 CONCLUSION

  • ScaleCast combines GRConv token mapping, GRAD windowed attention and gated residual updates, and increment decoding to integrate global guidance into autoregressive regional-state evolution.The model is trained on ERA5–CERRA pairs and reused with Pangu-Weather, GraphCast, and HRES without driver-specific retraining.Source: §5 · p. 9
  • GraphCast-driven ScaleCast improves 30-hour surface and upper-air forecasts and outperforms CERRA NWP on all nine variables.Windstorm and station case studies show improvements in cyclone positioning, core pressure, and selected local temperature and humidity changes.Source: §5 · p. 9
  • HRRR experiments demonstrate cross-dataset transferability, while evaluation across additional regions and forecasting systems is needed to establish broader applicability.Source: §5 · p. 9

AI USE STATEMENT

The statement describes the AI tools used during manuscript preparation and the authors’ responsibility for the paper’s scientific content.

AI USE STATEMENT

  • AI tools assisted with drafting, polishing, consistency checks, literature discovery, implementation inspection, and preparing analysis code and figures.Source: AI USE STATEMENT · p. 10
  • AI-generated text was not treated as scientific evidence or cited as a source.Source: AI USE STATEMENT · p. 10
  • The authors retain responsibility for verifying references and for the accuracy and integrity of the manuscript, including its methodology and empirical claims.Source: AI USE STATEMENT · p. 10

ETHICS STATEMENT

ScaleCast is intended for regional-weather-forecasting research, and its results should not be treated as operational certification or a replacement for established warning procedures.

ETHICS STATEMENT

  • ScaleCast is intended for research on regional weather forecasting.The study evaluates forecasts over Europe and cross-dataset transfer through HRRR fine-tuning, with additional station and windstorm verification.Source: ETHICS STATEMENT · p. 10
  • Forecast errors and dependence on upstream global guidance should be considered before operational use, particularly for severe-weather warnings and other safety-critical decisions.The passage does not characterize ScaleCast as operationally certified.Source: ETHICS STATEMENT · p. 10
  • The reported results do not constitute operational certification or replace established warning procedures.Source: ETHICS STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

The reproducibility statement maps the paper's sections and appendices to their methodological, experimental, and supporting materials.

REPRODUCIBILITY STATEMENT

  • Sections 3 and 4 describe the forecasting formulation, model components, datasets, baselines, and experimental settings.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Appendix A.1 specifies the reference architecture, with implementation and training details in following subsections.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Appendix B defines variables, error metrics, and aggregation procedures, while Appendix C provides lead-dependent results and additional comparisons and diagnostics.Appendix C includes global-driver comparisons, station and windstorm diagnostics, and design comparisons; data sources and baseline methods are identified through citations.Source: REPRODUCIBILITY STATEMENT · p. 10

A IMPLEMENTATION AND TRAINING DETAILS

The appendix records ScaleCast's reference configuration, transfer setup, token-mapping and decoder choices, and masked-learning adaptations. It also details HRRR fine-tuning and the limits of the adapted FAMO controller.

A.1 REFERENCE CONFIGURATION

  • The reference model uses six width-384 blocks with eight heads, 1,664 global tokens, 4,096 regional tokens, and regional windows of 4 × 4 tokens.The inputs use patch sizes four and eight; positional encodings identify patch indices rather than geographic coordinates, and terrain enters only the regional encoder.Source: §A.1 · p. 14

A.2 HRRR TRANSFER SETUP.

  • HRRR transfer fine-tunes the ERA5-to-CERRA checkpoint, replacing terrain and reinitializing four output heads while retaining the backbone and five compatible heads.The setup conditions each sample on HRRR at t and ERA5 analysis at t + 3 h to predict HRRR at t + 3 h; November–December 2022 is the evaluation period.Source: §A.2 · p. 14

A.3 TOKEN MAPPING, ORDERED GATHERING AND RESIDUAL PATHS

  • GRConv maps joint regional and global features to regional token positions, while windowed attention gathers nine windows' worth of key/value tokens in an ordered, not uniformly centered, sequence.The learned positional tensor is indexed by key-sequence position and added to keys; distinct inner and outer residual paths preserve regional features.Source: §A.3 · p. 14

A.4 VARIABLE-SPECIFIC INCREMENT DECODER

  • The decoder reshapes regional tokens into a patch field, applies a shared transposed-convolution decoder and variable-specific heads, then adds the decoded increment to the standardized regional input.The low-rank branch's final projection is zero-initialized, leaving the base decoder unchanged at initialization.Source: §A.4 · pp. 14–15

A.5 MASKED LEARNING AND THE ADAPTED FAMO CONTROLLER

  • Masked loss pools squared errors and valid-target counts over the effective batch, while the adapted FAMO controller uses lagged loss progress and sparse post-update probes.The implementation uses a 20-update warmup and probes every twentieth successful update; its adaptations mean the original FAMO guarantees do not automatically apply.Source: §A.5 · p. 15

B.1 VARIABLE INVENTORY AND REPORTED FIELD ERRORS

Frozen GraphCast and HRES drivers reduce several near-surface errors relative to Pangu, while frozen GFS input remains stable but loses accuracy across variables. Recursive trajectories and case diagnostics qualify these comparisons, and the section defines its error-score conventions.

B.1 VARIABLE INVENTORY AND REPORTED FIELD ERRORS

  • Per-lead pixel-pooled RMSE uses valid physical forecast errors, and valid grid cells receive equal weight; MAE and signed bias substitute absolute and signed errors, respectively.Pooled scores across leads combine squared errors and valid counts before taking the square root, rather than averaging lead-wise RMSE values; MSLP RMSE is reported in Pa.Source: §B.1 · p. 15, Table 3 · p. 16

C.1 FORECAST TRAJECTORIES AND GLOBAL-DRIVER SENSITIVITY

  • The ten recursive forecast leads test how the frozen regional predictor responds to its global driver and how errors evolve with lead time.The absolute-field case complements the error maps in the main text.Source: §C.1 · p. 16

C.1.1 FROZEN TRANSFER TO GRAPHCAST AND HRES FORCING

  • GraphCast and HRES reduce MSLP, T2m, and RH2m errors relative to Pangu-Weather, while upper-air performance varies by variable.This comparison substitutes global drivers without regional retraining; Figure 7 shows the differences by forecast lead.Source: §C.1.1 · p. 16, Figure 7 · p. 17

C.1.2 FROZEN TRANSFER FROM PANGU TO GFS FORCING

  • With frozen initialization and normalization, GFS forcing remains numerically stable but increases pooled error for all nine variables by 26.7%–98.6%.The increase ranges from 26.7% for RH850 to 98.6% for Z850; 12-hour GFS spacing prevents separating driver shift from temporal-interpolation effects, so this tests compatibility, not invariance or retraining.Source: §C.1.2 · p. 17, Table 4 · p. 17, Figure 8 · p. 18

C.1.3 COMPLETE 30-HOUR LEAD CURVES

  • ScaleCast, Hi-LAM, and StormCast-Reg remain comparatively stable across ten recursive steps, although humidity variables show substantial case-dependent oscillation.The passage cautions that pooled scalar scores should be read together with the trajectories; Figure 10 supplies per-lead MAE curves.Source: §C.1.3 · p. 18, Figure 9 · p. 19, Figure 10 · p. 20

C.1.4 ABSOLUTE REGIONAL FORECAST FIELDS

  • Figure 11 presents absolute T2m and RH2m fields on shared colour scales across the reference and forecasts, complementing the main-text error maps.Source: §C.1.4 · p. 20, Figure 11 · p. 21

C.2 WEATHER-SYSTEM AND STATION CASE STUDIES

  • The windstorm diagnostics include event 1812 track displacement and archived cases from February, April, and November 2022; archived event-1804 position values predate the main-text revision.The case-relative displacement plot uses along- and cross-track components at shared fixes; archived event comparisons are retained separately from the revised main-text comparison.Source: §C.2.1 · p. 21, Figure 12 · p. 21, Figure 13 · p. 22

C.2.1 WINDSTORM TRACKING ACROSS THE CANDIDATE EVENTS

  • The archived four-event windstorm figures report paired errors at shared times, coverage over eligible fixes, position error, and signed central-pressure error.Figure 15 also shows February tracking diagnostics for events 1804 and 1805; its shaded +66 h point is matched only by SC + GraphCast.Source: Figure 13 · p. 22, Figure 14 · p. 22, Figure 15 · p. 23

C.2.2 SPATIAL STATION-ERROR DIAGNOSTICS

The section documents HadISD spatial station-error panels and clarifies the scope of accompanying design comparisons. The supplied passages specify station masks and comparison settings but do not report forecast-error outcomes.

C.2.2 SPATIAL STATION-ERROR DIAGNOSTICS

  • All methods use the same valid station mask within each variable in the Figure 16 spatial panels.These panels accompany the station trajectories in Figure 5.Source: §C.2.2 · p. 23
  • Figure 16 maps forecast-minus-observation T2m errors at 1,626 stations and RH2m errors at 1,140 stations for a +9 h forecast.The initialization is 17 February 2022, 00 UTC.Source: Figure 16 · p. 23

C.3 DESIGN AND TRAINING COMPARISONS

  • One-step design-study scores use target-time global analyses, so they are diagnostics rather than direct rankings against the +30-hour forecast results.The main recursive comparison uses forecast trajectories.Source: §C.3 · p. 23
  • The design comparisons distinguish settings by controlled changes: cross-attention replaces regional attention, while direct prediction removes only the increment output.Geo-remap + fusion and GRConv share the regional backbone, training budget, and loss; Shared and Separate differ only in output heads.Source: §C.3 · p. 23

C.3.1 TRAINING STRATEGIES AND HUMIDITY OBJECTIVES

The training comparisons show trade-offs across variables, while the coarse-input comparison changes the surrounding global context without changing the regional prediction grid. The HRRR analysis figure specifies the fields and dates underlying regional comparisons.

C.3.1 TRAINING STRATEGIES AND HUMIDITY OBJECTIVES

  • RH2m RMSE falls from 5.336 to 5.253 percentage points, while MSLP error rises from 40.8 to 55.8 Pa and Z500 error from 0.476 to 0.626 dam.Aggregate nMSE rises from 3.009 to 3.038 in units of 10−2; the comparison uses a scale-and-terrain configuration with multiscale and gradient losses versus grouped optimization.Source: §C.3.1 · p. 24, Table 5 · p. 24
  • RH2m and RH500 RMSE fall from 5.283 to 5.176 and 8.196 to 8.121 percentage points, while aggregate nMSE falls from 2.974 to 2.922 in units of 10−2.MSLP, upper-air temperature, and geopotential errors increase; training budgets differ, so the selected checkpoints do not isolate the objective’s causal effect.Source: §C.3.1 · p. 24, Table 5 · p. 24

C.3.2 EXTENT OF THE GLOBAL INPUT

  • Expanding the coarse input from 128×128 to 128×208 adds surrounding atmospheric context, not a change to the predicted area or its 512×512 regional resolution.The smaller coarse domain already contains the full regional domain; Figure 17 shows the grid footprints and coordinate systems.Source: §C.3.2 · p. 24, Figure 17 · p. 24

T2m MSLMA

  • Figure 18 shows HRRR analysis fields for 09 UTC on 2 and 28 December 2022, with dates in rows and T2m and MSLMA in columns.The fields underlie the regional comparisons.Source: Figure 18 · p. 25
Loading 2610.12401v1…