PROSWIN: Probabilistic Solar Wind Speed Forecasting Using Deep Distributional Regression From Solar Images
Daniel Collin, Yuri Shprits, Luca Chiarabini, Stefan J. Hofmeister, Nadja Klein, Guillermo Gallego
astro-ph.SR · cs.LG · physics.data-an · physics.space-ph
Quick verdict
Claims first fully probabilistic ML solar-wind-speed forecast with <1% calibration deviation; evidence is strong, though HSS peaks still trail persistence.
Summary
PROSWIN predicts the hourly solar wind speed at Earth four days ahead as a full probability distribution rather than a single value. It couples a Swin Transformer image encoder (171 and 211 angstrom AIA channels) and physical features (49 lagged solar-wind speeds, solar-cycle indicators) to distributional regression that outputs log-normal parameters, trained in two stages. The authors also propose a 'prediction score', a weighted harmonic mean of timeline and HSS-peak errors, for model selection. On 14 years (June 2010-June 2024) under 5-fold cross-validation with 90-day gaps, the best model reaches a timeline CRPS of 41.0 km/s, HSS peak CRPS of 45.3 km/s and a prediction score of 42.3 km/s, with prediction-interval coverage deviating on average by only 0.9% (2.9% for peaks) from nominal. Against selected literature models it is the only one accurate on both timeline and peaks. On unseen data (July 2024-June 2026, solar maximum) it beats climatological and 27-day persistence baselines on timeline metrics, but peak accuracy degrades (peak CRPS 82.2 km/s) and persistence remains better on peaks. CMEs are not modeled, inputs use definitive rather than near-real-time data, and hyperparameter optimization is non-nested. Code, data and trained models are public.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 2
● low 4
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- PROSWIN is, to the authors' knowledge, the first fully probabilistic machine learning model for solar wind speed, forecasting calibrated log-normal distributions four days ahead.
- Predicted uncertainties are very well calibrated, with average prediction-interval coverage deviation below 1% on the timeline (0.9%) and 2.9% on HSS peaks.
- The proposed prediction score (weighted harmonic mean of timeline and HSS-peak errors) is an effective model-selection criterion that avoids trading timeline accuracy against peak accuracy.
- The 171 Å channel carries information complementary to the commonly used 193 Å and 211 Å channels, whose contents are largely redundant.
- Compared to selected literature models, PROSWIN is the only model accurate for both timeline and HSS peak values.
Strengths
First fully probabilistic ML model for solar wind speed, with rigorous validation design
5-fold CV with 20 nine-month chunks, 90-day gaps to block 27-day coronal-hole leakage, plus a genuine holdout window (Jul 2024-Jun 2026) untouched by training or hyperparameter search (Sections 4.2, 8).
Extensive input-channel ablation (22 model configurations)
Section 5 quantifies channel information: 193 and 211 Å largely redundant, 171 Å complementary, HMI alone a clear outlier with a plausible physical explanation (Section 5.1). Each configuration gets its own TPE hyperparameter optimization.
Calibration evidence is concrete and multi-faceted
Prediction score makes the timeline-vs-peak trade-off explicit and testable
Transparent self-critique and full reproducibility artifacts
Weaknesses
HSS peak accuracy loses to a naive 27-day persistence baseline
New prediction score metric has free weights and partial blind spots
Hyperparameter optimization is not nested
CMEs unmodeled with potential tail effects
Operational readiness untested
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
PROSWIN — Preliminary Review
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is PROSWIN and how does it forecast solar wind speed?
PROSWIN is a machine learning model that predicts the solar wind speed at Earth four days in advance as a full probability distribution instead of a single number. It feeds SDO solar images (171 and 211 angstrom channels), magnetogram-based preprocessing, and physical features (solar wind speeds from the previous solar rotation, sunspot numbers) into a Swin Transformer neural network whose heads output the parameters of a log-normal distribution, yielding prediction intervals and exceedance probabilities.
How accurate is the PROSWIN solar wind forecast?
On 14 years of cross-validated data (2010-2024) it reaches a timeline CRPS of about 41 km/s and an HSS peak CRPS of about 45 km/s, with prediction-interval coverage deviating under 1% from nominal. In a later solar-maximum period it had never seen (2024-2026), timeline accuracy stayed ahead of simple baselines, but peak errors grew (peak CRPS 82.2 km/s) and a naive 27-day persistence forecast remained more accurate on peak speeds.
What is the 'prediction score' introduced in this paper?
It is a new model-selection metric that combines two errors - accuracy over the whole timeline and accuracy on high-speed-stream peak values - using a weighted harmonic mean (weights 2:1). It was designed because standard average-error metrics reward models that neglect rare but impactful fast-wind peaks. It is meant for choosing between models rather than as a physical error measure.
Can PROSWIN predict coronal mass ejections (CMEs) and geomagnetic storms?
No. CMEs are not modeled and no CME information is provided as input; the authors show the model's uncertainty rises only ~3% during ICME disturbances, which they interpret as weak implicit signal. Since geomagnetic impact also depends on the IMF Bz component, PROSWIN should be seen as an ambient (HSS-driven) solar wind risk indicator, not a storm forecast.
Is the code and data for PROSWIN publicly available?
Yes. The paper cites a public Python implementation on Zenodo (doi 10.5281/zenodo.19630836) and the datasets plus trained models via GFZ Data Services (doi 10.5880/GFZ.OJSJ.2026.001), alongside published hyperparameter tables.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The extracted text omitted the full Tables 1-3 (main CV results and literature comparison) and parts of Sections 5.1, 7.1 and 10; conclusions for those sections rely on in-text numbers only.
- Figures 1-11 are not analyzable from text; calibration and time-series claims rest on the figure captions and quoted numbers.
- The embedded abstract ends mid-word ('These results'), consistent with extraction truncation; no substantive content loss identified beyond the tables/sections noted.
- This is a preprint under review at AGU Advances (arXiv:2609.26683), not a peer-reviewed published version.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.