Suwandee, Kanawut, Suwanwimolkul, Suwichaya · Oct 15, 2026
arXiv eess.SY paper proposes joint diffusion for sky-image irradiance forecasting
Markdown · JSON · Original source
Kanawut Suwandee and Suwichaya Suwanwimolkul propose forecasting future sky images and the clear-sky index together in one diffusion process. Their approach replaces a common two-stage design: one model predicts sky frames, then a separate regressor converts those frames into irradiance. The irradiance head instead reads a denoised preview of the future sky. The forecaster uses sky images and deterministic covariates, without measured irradiance at its input; the authors describe a site with a camera but no pyranometer as a possible setting. On 2023 SIRTA data, the authors report a pooled RMSE of 104.1 W m−2 and a CRPS skill score of +0.34 against smart persistence. They report higher CRPS skill than twelve baseline configurations and say it remained when retrained on 2021–2022 data. A separately trained regressor applied to the model’s generated frames had an RMSE of 150.6 W m−2, compared with 104.1 for the joint head. The authors describe the gap as information that joint training recovers and a cascade discards. Point accuracy was not best: a deterministic network that also reads measured irradiance reached 97.6 W m−2. These reported results concern the stated data; performance at other sites is not established by this abstract.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
eess.SY · Article
Source details
Original source date: Oct 15, 2026 (2026-10-15).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:23:37.529Z. First observed: 2026-10-16T04:30:53.969Z. Retrieved: 2026-11-08T20:17:36.903Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The evidence is the abstract; it reports results on 2023 SIRTA data and retraining on 2021–2022, but does not establish performance at other sites. The authors report a point-accuracy comparison with a network that reads measured irradiance, while the proposed forecaster does not take measured irradiance as input; these are different input conditions.
Open question
How well would the joint approach perform at a site with a camera but no pyranometer?
Source-stated
The authors propose jointly forecasting future sky images and clear-sky-index sequences in one diffusion process, rather than predicting frames and then applying a separately trained regressor.
We propose a single-stage alternative that co-denoises the future sky-image and clear-sky-index sequences in one diffusion process
Source-stated
The proposed forecaster uses sky images and deterministic covariates, without measured irradiance at its input; the authors identify camera-equipped sites without a pyranometer as a possible setting.
with no measured irradiance at its input --- the situation at a site with a camera but no pyranometer.
Source-stated
On 2023 SIRTA data, the authors report a pooled RMSE of 104.1 W m−2 and a CRPS skill score of +0.34 against smart persistence.
On 2023 SIRTA data the model reaches a pooled RMSE of $104.1$~W\,m$^{-2}$ and a skill score of $+0.34$ on the continuous ranked probability score (CRPS) against smart persistence
Source-stated
The authors report that their CRPS skill was higher than all twelve baseline configurations and remained when retrained on 2021–2022 data.
higher than any of twelve baseline configurations, and retains that skill when retrained on 2021--2022.
Source-stated
A separately trained regressor on the generated frames had RMSE 150.6 W m−2, versus 104.1 for the joint head; the authors describe the difference as recovered by joint training and discarded by a cascade.
a separately trained regressor reaches $150.6$~W\,m$^{-2}$ against the joint head's $104.1$: roughly $46$~W\,m$^{-2}$ that joint training recovers and a cascade discards.
Source-stated
The reported point accuracy was not best: a deterministic network that also reads measured irradiance attained 97.6 W m−2.
Point accuracy is competitive but not best: a deterministic network that also reads measured irradiance attains $97.6$~W\,m$^{-2}$.
Source and contributor context
Suwandee, Kanawut, Suwanwimolkul, Suwichaya · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6f9f545f-9740-4bee-8eef-81e367cfc42e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:23:37.529Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 37997052449d11b6891cb21ba1624f7b2871d6a3bff9923d94b4d121b7b88357. Source text hash: 617d0cc98f76d6a29e6d38e70d33cc7679b1036b45e8da349e3fc2be9fbb8351.
Revision and source-check history
Version 1 · 2026-11-08T20:23:37.529Z
arXiv eess.SY paper proposes joint diffusion for sky-image irradiance forecasting
Kanawut Suwandee and Suwichaya Suwanwimolkul propose forecasting future sky images and the clear-sky index together in one diffusion process. Their approach replaces a common two-stage design: one model predicts sky frames, then a separate regressor converts those frames into irradiance. The irradiance head instead reads a denoised preview of the future sky. The forecaster uses sky images and deterministic covariates, without measured irradiance at its input; the authors describe a site with a camera but no pyranometer as a possible setting. On 2023 SIRTA data, the authors report a pooled RMSE of 104.1 W m−2 and a CRPS skill score of +0.34 against smart persistence. They report higher CRPS skill than twelve baseline configurations and say it remained when retrained on 2021–2022 data. A separately trained regressor applied to the model’s generated frames had an RMSE of 150.6 W m−2, compared with 104.1 for the joint head. The authors describe the gap as information that joint training recovers and a cascade discards. Point accuracy was not best: a deterministic network that also reads measured irradiance reached 97.6 W m−2. These reported results concern the stated data; performance at other sites is not established by this abstract.
Original source
eess.SY · Article
Source details
Original source date: Oct 15, 2026 (2026-10-15).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-08T20:23:37.529Z. First observed: 2026-10-16T04:30:53.969Z. Retrieved: 2026-11-08T20:17:36.903Z.
Limits
The evidence is the abstract; it reports results on 2023 SIRTA data and retraining on 2021–2022, but does not establish performance at other sites. The authors report a point-accuracy comparison with a network that reads measured irradiance, while the proposed forecaster does not take measured irradiance as input; these are different input conditions.
Open question
How well would the joint approach perform at a site with a camera but no pyranometer?
Source-stated
The authors propose jointly forecasting future sky images and clear-sky-index sequences in one diffusion process, rather than predicting frames and then applying a separately trained regressor.
We propose a single-stage alternative that co-denoises the future sky-image and clear-sky-index sequences in one diffusion process
Source-stated
The proposed forecaster uses sky images and deterministic covariates, without measured irradiance at its input; the authors identify camera-equipped sites without a pyranometer as a possible setting.
with no measured irradiance at its input --- the situation at a site with a camera but no pyranometer.
Source-stated
On 2023 SIRTA data, the authors report a pooled RMSE of 104.1 W m−2 and a CRPS skill score of +0.34 against smart persistence.
On 2023 SIRTA data the model reaches a pooled RMSE of $104.1$~W\,m$^{-2}$ and a skill score of $+0.34$ on the continuous ranked probability score (CRPS) against smart persistence
Source-stated
The authors report that their CRPS skill was higher than all twelve baseline configurations and remained when retrained on 2021–2022 data.
higher than any of twelve baseline configurations, and retains that skill when retrained on 2021--2022.
Source-stated
A separately trained regressor on the generated frames had RMSE 150.6 W m−2, versus 104.1 for the joint head; the authors describe the difference as recovered by joint training and discarded by a cascade.
a separately trained regressor reaches $150.6$~W\,m$^{-2}$ against the joint head's $104.1$: roughly $46$~W\,m$^{-2}$ that joint training recovers and a cascade discards.
Source-stated
The reported point accuracy was not best: a deterministic network that also reads measured irradiance attained 97.6 W m−2.
Point accuracy is competitive but not best: a deterministic network that also reads measured irradiance attains $97.6$~W\,m$^{-2}$.
Source and contributor context
Suwandee, Kanawut, Suwanwimolkul, Suwichaya · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 6f9f545f-9740-4bee-8eef-81e367cfc42e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-08T20:23:37.529Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 37997052449d11b6891cb21ba1624f7b2871d6a3bff9923d94b4d121b7b88357. Source text hash: 617d0cc98f76d6a29e6d38e70d33cc7679b1036b45e8da349e3fc2be9fbb8351.
uncertain · 2026-11-08T20:23:37.529Z · anthropic/claude-opus-5.5
Publication: 6f9f545f-9740-4bee-8eef-81e367cfc42e
attribution — pass: The reading correctly credits Kanawut Suwandee and Suwichaya Suwanwimolkul and keeps results framed as author reports. The title's 'arXiv eess.SY' label matches the feed publisher, although the primary listing is cs.CV with an eess.SY cross-list. This is not material.
Authors:Kanawut Suwandee, Suwichaya Suwanwimolkul
summary — uncertain: Central terms are left unexplained for a non-specialist reader: irradiance, clear-sky index, diffusion process, pyranometer, RMSE, CRPS and smart persistence. The reading also never says what such forecasts are used for. Smallest repair: add tooltips or a one-line plain gloss for irradiance and pyranometer. The summary also omits the authors' counterweight to 'not best' point accuracy: 'competitive', and a CRPS lead of 0.26 to 0.30 in every sky-condition tertile. Leaving these out understates the paper's argument; restore that sentence.
The model's CRPS skill leads that network's in every sky-condition tertile, by $0.26$ to $0.30$.
date — pass: The original publication date of 2026-10-15 matches the v1 submission. The 5 Nov 2026 date is a later v2 revision, and the feed date is later still.
Submitted on 15 Oct 2026 (v1), last revised 5 Nov 2026 (this version, v2)
support — pass: Each claim tracks the abstract with the authors' figures: 104.1, +0.34, twelve baselines, the 2021–2022 retraining, 150.6 vs 104.1, and 97.6. The limitations about different input conditions and other sites are fair. The question follows from the stated no-pyranometer use case and is not a leading claim.
Point accuracy is competitive but not best: a deterministic network that also reads measured irradiance attains $97.6$~W\,m$^{-2}$.
media — pass: The selected media is the verified arXiv abstract page, which is the direct original source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv