Source-linked AI summary
A comparative study of convolutional neural network models for wind field downscaling
Kevin Höhlein, Michael Kern, Timothy Hewson, Rüdiger Westermann
TL;DR
Fine-scale wind prediction over complex terrain requires information that coarse simulations may not resolve, and the added value of nonlinear downscaling over classical methods has remained uncertain. The study compares CNNs and multilinear regression for downscaling 31 km ERA5 forecasts to 9 km HRES fields and proposes DeepRU. DeepRU outperforms the linear model and EnhanceNet in reconstruction accuracy, while deep CNNs better reproduce local flow variance in the studied settings.
Problem
Coarse simulations may not resolve topographically driven local winds, and whether nonlinear machine-learning methods add value over classical downscaling remains unsettled.
Source: §1 · p. 2Method
The study compares CNN architectures with multilinear regression for downscaling 31 km ERA5 short-range forecasts to 9 km HRES wind fields, and proposes a residual U-Net called DeepRU.
Source: §3.2 · pp. 5–6, §8 · p. 24Results
DeepRU outperforms the linear model and EnhanceNet in reconstruction accuracy, while deep CNNs better replicate local-flow variance, especially in mountainous areas and coastal seas.
Source: §8 · pp. 24–25Takeaways & Limitations
Deep CNN downscaling is particularly suited to high-magnification ratios over large domains, where linear link functions can be insufficient for local flow variability.
Source: §8 · p. 25The study uses 100 m winds because ECMWF output apparently represents 10 m winds poorly over mountains.
Source: §7 · p. 21
Abstract
from arXiv · showhide
We analyze the applicability of convolutional neural network (CNN) architectures for downscaling of short-range forecasts of near-surface winds on extended spatial domains. Short-range wind field forecasts (at the 100 m level) from ECMWF ERA5 reanalysis initial conditions at 31 km horizontal resolution are downscaled to mimic HRES (deterministic) short-range forecasts at 9 km resolution. We evaluate the downscaling quality of four exemplary model architectures and compare these against a multi-linear regression model. We conduct a qualitative and quantitative comparison of model predictions and examine whether the predictive skill of CNNs can be enhanced by incorporating additional atmospheric variables, such as geopotential height and forecast surface roughness, or static high-resolution fields, like land-sea mask and topography. We further propose DeepRU, a novel U-Net-based CNN architecture, which is able to infer situation-dependent wind structures that cannot be reconstructed by other models. Inferring a target 9 km resolution wind field from the low-resolution input fields over the Alpine area takes less than 10 milliseconds on our GPU target architecture, which compares favorably to an overhead in simulation time of minutes or hours between low- and high-resolution forecast simulations.
1 | INTRODUCTION AND CONTRIBUTION
Fine-scale wind forecasts are costly, while coarse models can miss topographically driven local wind structures and the value of nonlinear statistical methods over classical approaches remains unsettled. This study compares CNN architectures with multilinear regression and proposes DeepRU, finding that appropriate architecture design and tuning can reproduce local- and global-scale structures with high fidelity.
1 | INTRODUCTION AND CONTRIBUTION
- High-resolution simulations are needed for reliable low-level winds in complex terrain, but dynamical downscaling remains computationally demanding and time-consuming.Sub-grid features such as slopes, valleys, ridges, and cliffs can shape local winds in ways that global models cannot resolve.Source: §1 · p. 2
- Whether nonlinear machine-learning methods add value over classical statistical approaches remains unresolved in prior downscaling work.Statistical downscaling learns relationships between coarse predictor data and fine-scale predictand data from paired examples.Source: §1 · p. 2
1.1 | Contribution
- DeepRU, a U-Net-based CNN, improves reconstruction quality over existing architectures through design patterns and training adaptations.The study compares multilinear regression, a lightweight linear convolutional model, and increasingly complex nonlinear convolutional models.Source: §1.1 · p. 3
- The study tests whether additional climate variables and high-resolution topography improve inference, and trains on subregions to limit reliance on geographic location.It also evaluates relationships between model complexity, prediction quality, memory use, and prediction time.Source: §1.1 · p. 3
- Appropriate architecture design and tuning enable network-based downscaling to reproduce local- and global-scale wind structures with high fidelity.The study compares reconstruction quality with high-resolution simulations of real-world weather over complex terrain in central and southern Europe.Source: §1.1 · p. 3
2 | RELATED WORK
Prior empirical-statistical downscaling work spans regression, weather-typing, and stochastic approaches, but reproducing fine-scale spatial variability remains difficult and findings on machine learning’s added value vary. CNNs have also been used for image and scientific-field upscaling, providing related approaches to reconstructing high-resolution data.
2.1 | Empirical-statistical downscaling
- Empirical-statistical downscaling includes regression, weather typing, and stochastic weather generators, while realistic fine-scale spatial variability remains challenging for many models.The figure illustrates low-resolution, HRES, and DeepRU wind fields for a coastal region and the Swiss Alps.Source: §2.1 · pp. 3–4, Figure 1 · p. 4
- Comparisons of non-classical machine learning with classical downscaling methods have produced mixed findings, leaving the useful conditions for deep learning uncertain.Earlier CNN downscaling studies largely addressed precipitation or temperature and did not use newer residual, skip-connection, or U-Net patterns.Source: §2.1 · p. 4
2.2 | Upscaling of images and physical fields
- CNNs have supported image super-resolution and reconstruction of scientific scalar and vector fields, including volumetric and time-varying data.In image super-resolution, the goal is to increase resolution while maintaining visual quality and avoiding artifacts and blurriness.Source: §2.2 · pp. 4–5
3 | TRAINING DATA
The study uses ECMWF short-range forecast fields over a topographically varied European domain, mapping irregular low- and high-resolution grids to regular arrays for CNN processing. Inputs include wind and selected supplementary predictors; padding and scaling prepare the data for training.
3 | TRAINING DATA
- Training and evaluation use ECMWF short-range forecast data for near-surface winds over central and southern Europe.The domain spans mountains, flatter areas, Mediterranean waters, and coastal regions where topography can influence wind and low- and high-resolution simulations can differ.Source: §3 · p. 5, §3.1 · p. 5
3.2 | Low- and high-resolution simulations
- The low-resolution ERA5 inputs are 31 km hourly forecasts, and the HRES targets are 9 km forecasts selected at lead times of 7–18 hours.The analysis covers March 2016 to October 2019, when HRES O1280 products were available.Source: §3.2 · pp. 5–6
3.3 | Predictor and predictand variables
- Wind components at 100 m are the main predictors and targets, with altitude and land-sea mask as static predictors and BLH, FSR, and Z500 as additional dynamic predictors.BLH represents the lowest atmospheric layer’s vertical extent; FSR reflects frictional retardation, while Z500 represents the 500 hPa pressure-level height and synoptic flow.Source: §3.3 · p. 6
3.4 | Data padding
- Irregular Gaussian-grid data are padded into regular arrays rather than interpolated, preserving local detail and spatial adjacency for CNN processing.Boundary values are replicated, and a binary mask distinguishes valid from padded cells in the loss computation.Source: §3.4 · pp. 6–7
3.5 | Data scaling
- Predictors are standardized before training, using node-wise means and standard deviations for time-varying fields and domain statistics for static fields.Predictands are standardized as well, which the authors found useful given differences in average wind speeds between coastal or sea sites and mountain ranges.Source: §3.5 · p. 7
4 | NETWORK ARCHITECTURES
The section compares CNN architectures for mapping coarse predictors to fine-resolution winds against a localized multilinear-regression baseline, and proposes DeepRU to capture wind structures across spatial scales.
4.1 | Linear convolutional network model: LinearCNN
- LinearCNN learns local linear mappings through separate low- and high-resolution branches, expanding low-resolution features with a transposed convolution to estimate each target region.Its low-resolution branch encodes a 5 × 5 input patch and decodes it to a 12 × 9 high-resolution estimate; the high-resolution branch adds a local correction.Source: §4.1 · pp. 8–9
4.2 | Simple nonlinear CNN: DeepSD
- DeepSD adapts the three-layer SRCNN pipeline, using one stage for the 3× longitude and 4× latitude magnification and adding high-resolution orography to its inputs.The low-resolution predictors are bicubic-interpolated before being concatenated with high-resolution predictors.Source: §4.2 · p. 9
4.3 | Fast nonlinear CNN: FSRCNN
- FSRCNN processes low-resolution inputs with smaller kernels, then uses staged bilinear upsampling instead of transpose convolution; batch normalization stabilized training of this deeper variant.The staged upsampling replaces one large transpose convolution, which the authors found could slow training or prevent convergence.Source: §4.3 · pp. 9–10
4.4 | Deep nonlinear CNN: EnhanceNet
- EnhanceNet uses 20 processing convolutions in residual blocks and adds its output to a bicubic-interpolated wind-field baseline, with pixel-wise losses for the accuracy-focused task.Its 20 layers are arranged as 10 two-layer residual blocks.Source: §4.4 · p. 11
4.5 | DeepRU
- DeepRU is a six-stage residual U-Net with multiscale skip connections; residual blocks and high-/low-resolution skip connections with input interpolation improved reconstruction accuracy.The architecture concatenates encoding and decoding features to preserve detail, and adds its output to a bicubic wind interpolant.Source: §4.5 · pp. 12–13
4.6 | Localized multi-linear regression model: LinearEnsemble
- LinearEnsemble predicts each high-resolution wind node from its k nearest low-resolution nodes, imposing sparsity on the regression weights to limit parameter growth with domain size.The tested variants use either two wind-component predictors or five predictors including BLH, FSR, and Z500; neighborhoods beyond k = 16 did not significantly improve results.Source: §4.6 · p. 13
5 | TRAINING METHODOLOGY
Training uses cross-validation and randomly sampled sub-patches to address limited, temporally correlated data; losses and regularization are chosen to support wind-field reconstruction and stable optimization.
5.1 | Cross-validation
- Three validation cycles use disjoint consecutive full-year intervals, reducing temporal information overlap and seasonality effects in performance estimates.Each cycle trains three random initializations and selects the model with the lowest average validation loss.Source: §5.1 · pp. 13–14
5.2 | Patch training
- Intermediate patch sizes yielded the best training results, while very small patches reduced prediction quality, possibly because they omit necessary context.Patch training also limits inputs to local information, reducing overfitting to statistical artifacts and potentially non-physical long-distance correlations.Source: §5.2 · p. 14
5.3 | Loss functions
- MSE is invariant to local wind-direction rotations but weights deviations quadratically by wind magnitude; node-wise target standardization reduces spatial inhomogeneity before training.Predictions and targets are converted back to their real scale for validation and evaluation.Source: §5.3 · pp. 14–15
5.4 | Implementation and optimization
- ADAM optimization begins at a learning rate of 10^-3, reduces it on validation-loss plateaus, and continues for 150 epochs per run without early stopping.Training and validation losses usually saturated after 50–60 epochs, with only minor variation afterward.Source: §5.4 · p. 15
5.5 | Regularization
- Weight decay is applied to all models, while nonlinear CNNs use batch normalization and DeepRU additionally uses 2D dropout at rate 0.1 after each residual block.The authors report that batch normalization accelerated convergence and that they observed no negative effects from combining batch normalization and dropout.Source: §5.5 · p. 15
6 | EVALUATION
The evaluation compares CNN configurations and predictor sets using validation errors, spatial patterns, feature importance, and example flow reconstructions. It also examines how DeepRU and LinearEnsemble represent flow structures in two analyzed cases.
6 | EVALUATION
- CNN predictor configurations varied low- and high-resolution orography and dynamic inputs; DeepSD was evaluated only with configurations (B) and (D), and LinearEnsemble only with (A) and (C).LinearEnsemble excluded static predictors, whereas DeepSD retained high-resolution orography to stay close to its original implementation.Source: §6 · p. 15, Table 1 · p. 16
6.1 | Run-time performance and memory requirements
- DeepRU and EnhanceNet used more memory as model complexity increased, yet nonlinear models achieved the best overall results; DeepRU had the longest training and prediction times among the models.LinearEnsemble required roughly 137 GiB for its three-year training dataset, so data were streamed from disk, yielding the highest training time among trained models.Source: §6.1 · p. 16
6.2 | Quantitative analysis
- DeepRU configuration (D) was the most accurate model, with an average MSE of around 2.7 (m s–1)2, and it reduced MSE beyond LinearEnsemble while benefiting from additional predictors.EnhanceNet first surpassed LinearEnsemble when high-resolution orography predictors were included; LinearEnsemble variants with more predictors tended to overfit.Source: Figure 8 · p. 17, §6.2 · p. 17
6.3 | Spatial distribution of prediction errors
- DeepRU showed the lowest cosine and magnitude differences, captured local wind-speed and direction variance better than LinearEnsemble, and retained smaller spatial error amplitudes.LinearEnsemble tended to underestimate wind magnitudes and missed extreme shifts near mountains; DeepRU also showed wave-like magnitude-error patterns near Mediterranean topographic boundaries.Source: Figure 9 · p. 18, §6.3 · p. 18
6.4 | Analysis of feature importance
- Wind components U and V had the largest perturbation effects across architectures, while FSR and BLH had little impact and Z500 mattered mainly to LinearCNN and DeepSD.Change ratios should be interpreted within each model, since more complex models had larger ratios; LinearEnsemble alone showed correlations between FSR and high-resolution winds.Source: §6.4 · p. 19
6.5 | Analysis of reconstructed flow patterns
- In high-inhomogeneity cases, DeepRU used local and global orographic information to reproduce HRES flow features that LinearEnsemble missed, while both models handled Austrian Alps conditions reasonably well.DeepRU better captured small-scale magnitude transitions near the Alps and orography-mediated flow adjustments in eastern France; the evaluation compared the best nonlinear model with the localized linear model.Source: §6.5 · pp. 20–21
- In case A, DeepRU better mirrors HRES by elongating Ligurian vortex flow toward northern Italy; in case B, it captures sharp HRES wind-magnitude changes that LinearEnsemble misses.In case A, DeepRU also enhances southerly wind near Corsica’s western coast. In case B, high winds interleaved with topographically triggered vortices occur along the Apennines and near three marked lakes.Source: §6.5 · p. 21
7 | APPLICATIONS IN FORECASTING
The authors present the CNN approach as a blueprint for future forecasting systems, not a direct operational replacement at the study’s resolutions. They outline finer-resolution training data and describe possible uses across forecast settings, while noting boundaries in wind-height representation and the study’s scope.
7 | APPLICATIONS IN FORECASTING
- Predictors at 5–10 km and training targets at 1–2 km are proposed as a route toward future operational use, since the study’s resolutions are not competitive with current operational configurations.The finer-resolution target data would be needed for training, not real-time operation.Source: §7 · p. 21
- Where limited-area models (LAMs) are unavailable, CNN predictions could cover short- and medium-range leads; where LAMs exist, use could focus on the medium range.Training on the same LAM could provide continuity across the LAM–global temporal boundary.Source: §7 · p. 21
- ECMWF’s poor representation of 10 m winds over mountains motivates using 100 m winds; a 100 m predictor could target 10 m winds if the latter are better represented at 1–2 km.The authors note some evidence for improved 10 m representation at that resolution in limited-area models.Source: §7 · pp. 21–22
- Renewables and other applications needing mean speeds across the full speed range may benefit, but extreme-wind predictions need further investigation because the training period may be insufficient.The authors warn that predictions may be systematically too weak or unstable.Source: §7 · p. 22
- The study did not directly explore gusts, including extreme gusts, which are important because of their risks to life and infrastructure.Source: §7 · p. 22
- The study’s region, relatively remote from storm tracks, may help explain why boundary-layer height added little predictive value for the CNNs.The authors note that low-level stability matters for cyclone-related gusts, so this finding may not carry over to storm-track regions.Source: §7 · p. 23
8 | DISCUSSION AND OUTLOOK
DeepRU’s multiscale architecture outperforms the linear model and EnhanceNet in reconstruction accuracy, but model comparisons are incomplete and performance depends on domain and temporal setting. The discussion also identifies geometry, broader climate variables, and finer-scale or probabilistic forecasting as directions for further work.
8 | DISCUSSION AND OUTLOOK
- The architecture search was limited: training costs precluded complete direct comparison, and alternative architectures may match or exceed the results with lower complexity.The proposed architecture is the endpoint of iterative training, evaluation, and refinement across a limited set of design patterns.Source: §8 · p. 23
- The linear ensemble is more accurate than shallow nonlinear CNNs, while deeper CNNs compete with it and can surpass it as predictors and high-resolution topography are added.EnhanceNet improves with additional predictors, whereas LinearEnsemble cannot use them and tends to overfit.Source: §8 · pp. 23–24
- DeepRU outperforms both the linear model and EnhanceNet in reconstruction accuracy, using features at different spatial scales to align local wind systems with large-scale flow patterns.The authors report remaining deviations from native high-resolution forecasts.Source: §8 · p. 24
- The study focuses on low-level winds; other climate variables may require different approaches, while suitable predictors could let these architectures support other variables.The authors note that precipitation may require a probabilistic approach because high-resolution outcomes can vary substantially for a given low-resolution representation.Source: §8 · pp. 25–26
- Applying trained networks to North American winds reduced performance by about 70% in MAE and 90% in cosine dissimilarity, while training on North American data achieved similar reconstruction quality.The authors therefore advise careful assessment of generalizability across domains.Source: §8 · p. 26
- Temporal coherence was not enforced: training focused on spatial correlations at single time steps, and the authors propose modeling temporal correlations across multiple steps.They also note that temporal correlation measures for vector-valued winds are non-trivial to define.Source: §8 · p. 26
- CNN treatment of spherical geometry may become problematic on larger domains, where disregarding Earth’s curvature can produce distortion artifacts.The authors suggest spherical CNNs or geometric deep learning as possible ways to address this limitation.Source: §8 · pp. 26–27
- The authors envision operational downscaling to 1–2 km and probabilistic ensemble outputs, where generative CNNs could represent ambiguity among fine-scale realizations.These are proposed future applications, not evaluated outcomes of this study.Source: §8 · p. 27
9 | CONCLUSION
The study compares CNN complexity levels for extended-domain wind downscaling against localized linear regression models. It finds that deeper nonlinear models exploit nonlinear relationships and benefit from high-resolution static predictors, while additional predefined low-resolution dynamic predictors provide only minor improvements.
9 | CONCLUSION
- The study examines CNNs for extended-domain wind downscaling, varying network complexity and comparing the models with an ensemble of localized linear regression models.Source: §9 · p. 27
- Deeper, more complex CNNs exploit nonlinear correlations to map between low- and high-resolution fields, and all nonlinear models benefit from high-resolution static predictors such as local orography.The three predefined low-resolution dynamic predictors provide only minor improvements.Source: §9 · p. 27
- The authors identify temporal information and generative neural networks for probabilistic downscaling as future research directions.They argue that the demonstrated CNN performance should motivate further research on these architectures for predictive tasks.Source: §9 · p. 27