Transformers Applied to Short-term Solar PV Power Output Forecasting

Andea Scott; Folasade Ayoola; Sindhu Sreedhara

REVIEW 4 major objections 7 minor 13 references

Transformers Applied to Short-term Solar PV Power Output Forecasting

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A frozen vision transformer nearly matches a CNN baseline for short-term solar PV power forecasting, though it lags on sunny days; the paper's headline claim is that a ViT predicts solar output from sky images almost as well as the SUNSET…

desk verdict The abstract claims the ViT is almost as good as the CNN baseline, but the paper's own tables show it is 42% worse overall and five times worse on sunny days, so the central claim is contradicted by the reported numbers. read the letter →

arxiv 2505.03188 v1 pith:4BNJU3EH submitted 2025-05-06 cs.CE

Andea Scott , Sindhu Sreedhara , Folasade Ayoola This is my paper · ORCID

classification cs.CE

keywords solarPVforecastingvisiontransformerskyimagesnowcastingtransferlearningregressionrenewableenergy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether a pure transformer architecture can replace a convolutional neural network in the task of predicting solar PV power output from sky images. Using a frozen, pretrained HuggingFace Vision Transformer with a small trainable head, the authors report RMSE values close to those of the SUNSET CNN baseline on a held-out test set, concluding that the transformer performs almost as well in the PV output prediction task. The significance, if the claim holds, is that a generic image-classification transformer pretrained on ImageNet can serve as a near-parity feature extractor for a specialized renewable-energy regression problem without bespoke architecture design. The paper also notes that the transformer performs worse on sunny days, a caveat that matters because sunny conditions are precisely where solar forecasting is easiest and most reliable.

What carries the argument

The central object is the HuggingFace Vision Transformer (ViT), a pure transformer architecture that splits an image into 16x16 patches, embeds them linearly, and processes the sequence with self-attention layers. The paper loads weights pretrained on ImageNet, freezes all but the final layer normalization, pooler, and fully connected head, and trains only that head with MSE loss to regress PV power output. The load-bearing mechanism is feature transfer: the attention-based representations learned on natural images are reused, without fine-tuning the body, as the input to a small regression head for solar forecasting.

What would settle it

Re-run the SUNSET CNN baseline on the identical 20-day test set, using the same 224x224 rescaling and the same train/validation split protocol applied to the ViT, and compare RMSEs. If the baseline's overall RMSE rises from 2.36 kW to roughly the ViT's 3.35 kW or higher, the 'almost as well' conclusion is a pipeline artifact; if it stays near 2.36 kW, the ViT is genuinely worse than the CNN on this task.

Watch

Extended reading notes

Core claim

A pretrained Vision Transformer (ViT) with its weights mostly frozen, followed by a fully-connected regression head, can predict solar PV power output from a sky image with root mean squared error (RMSE) of 3.35 kW overall on a 20-day test set, compared with the SUNSET CNN baseline's 2.36 kW. On cloudy days the ViT reaches 4.01 kW versus the baseline's 3.38 kW, and on sunny days 2.62 kW versus 0.52 kW. The paper states this means the transformer performs almost as well as the baseline in the PV output prediction task, while acknowledging it performs worse on sunny days.

Load-bearing premise

The comparison rests on assuming the baseline SUNSET RMSE numbers are commensurable with the ViT numbers, even though the baseline was never re-run in this pipeline and the two models use different image resolutions, preprocessing, and train/validation splits.

Editorial extensions

If this is right

If the ViT nearly matches the CNN baseline, then future solar nowcasting systems can be built from off-the-shelf pretrained image transformers, saving the effort of designing and training task-specific CNNs from scratch.
The sunny-day degradation, if real, would push hybrid designs that switch to a simpler model or a different head when sky conditions are clear.
The ViT's ability to work as a feature extractor at 224x224 resolution suggests resolution flexibility: the model can ingest lower-resolution sky images and still produce usable forecasts, which is relevant for deployment on constrained hardware.
Since the model is frozen, the result would support a workflow where the expensive transformer body is shared across many solar sites and only the small regression head is retrained per site.
The extension of the same architecture to the SUNSET Forecast problem (multi-step-ahead forecasting) becomes a low-cost next step if the nowcast result is taken at face value.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's own tables show the ViT is worse on both sunny and cloudy days, and roughly 42% worse overall, so the phrase 'almost as well' is doing more work than the reported RMSEs support; a fair reader should treat the headline claim as optimistic relative to the numbers.
Because the baseline numbers come from prior work measured at 64x64 resolution while the ViT pipeline rescales to 224x224 inside the HuggingFace feature extractor, the comparison likely bundles preprocessing differences into the architecture comparison; an apples-to-apples rerun could shift both the magnitude and the sign of the gap.
The near-parity on cloudy days, where the ViT error is 4.01 versus 3.38, suggests the transformer captures cloud-structure cues reasonably well even without training the body, which if confirmed would make the attention mechanism a genuinely competitive feature extractor for irradiance-driven tasks.
A testable extension is to evaluate the frozen ViT on the same 64x64 inputs the baseline actually used, by downscaling inside the model or retraining the head on 64x64 patches, to separate the architecture effect from the resolution effect.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Thanks for the report. I read the paper myself and I agree with the rejection. The abstract says the transformer "performs almost as well as the baseline" and is worse only on sunny days. Table 1 gives SUNSET RMSEs of 0.52 sunny, 3.38 cloudy, 2.36 overall; Table 5 gives ViT 2.62, 4.01, 3.35. That is five times worse on sunny days and 42% worse overall. Cloudy is worse too. That is not "almost as well" by any reading.

Credit where due: the paper does one genuinely new empirical thing. It applies a frozen pretrained HuggingFace ViT to the SUNSET Nowcast test set and reports the resulting RMSEs. That specific measurement does not appear in any cited paper, and the authors are transparent enough to put the numbers in tables rather than hiding them. The data collection effort (103k sky-image/power pairs) is real, and using an external baseline rather than inventing their own is the right instinct.

But the soft spots are load-bearing. The RMSE numbers do not stay consistent: 2.61 in the pilot, 3.38 in Table 2, 2.61 in Table 4, 3.35 in Table 5. Table 4 literally has XX placeholders. The test set is described as 20 days in one place and ten days in another. The conclusion is truncated mid-sentence. No code, data, or seeds are provided. The baseline SUNSET numbers are imported from prior work, not recomputed in the ViT pipeline, so the comparison may be a preprocessing artifact: the ViT goes through a 224x224 feature extractor while the baseline was evaluated at 64x64. That does not explain the internal contradictions, but it means even the fixed numbers would not be commensurable without extra work.

The citation pattern is thin — no transformer-based solar forecasting work after 2020 — but that is a minor issue compared to the claim/data mismatch.

My assessment: this is an unfinished project draft, not a paper. The central claim is contradicted by its own evidence, and the completeness problems (XX, truncated text, contradictory test-set counts) would make referee time a waste. I would recommend desk rejection, not because the idea is bad — applying a ViT to sky-image regression is a reasonable thing to try — but because the authors need to first produce a consistent set of numbers and a properly rerun baseline. If they do that, the single negative data point for a frozen pretrained ViT might be worth a short note in a specialist venue. As submitted, it is not ready for peer review.

Referee Report

4 major / 7 minor

Summary. The paper applies a pretrained HuggingFace Vision Transformer (ViT) with a fully-connected regression head to the task of short-term solar PV power output nowcasting from sky images, comparing its RMSE with the CNN-based SUNSET baseline of the Stanford EAO lab. The authors report experiments with a 654-pair pilot, a warm-start full-data run, and a hyperparameter-tuned frozen-layer run, and they claim in the abstract that the transformer performs almost as well as the baseline but worse on sunny days.

Significance. If the central claim were supported, the paper would show that a frozen, pretrained ViT is a near-parity substitute for a task-specific CNN in a regression-from-sky-images problem, which would be a useful data point for the solar-forecasting community. The paper also leverages publicly available pretrained weights and a real-world dataset, which is a strength. However, the manuscript does not deliver a validated comparison: the baseline numbers are imported rather than recomputed, the ViT results are internally inconsistent, and the abstract's parity claim is contradicted by the paper's own tables. These problems are load-bearing rather than cosmetic, so the result as presented cannot be accepted.

major comments (4)

[Abstract; Section 5.3, Table 5 vs Section 4.1, Table 1] The abstract's central claim that the transformer 'performs almost as well as the baseline model' and 'performs worse on sunny days' is contradicted by the reported quantitative results. Table 5 gives ViT RMSE values of 2.62 (sunny), 4.01 (cloudy), and 3.35 (overall), while Table 1 gives SUNSET baseline RMSE values of 0.52 (sunny), 3.38 (cloudy), and 2.36 (overall). On these numbers the ViT is worse in every category: about five times worse on sunny days and about 42% worse overall, and it is also worse on cloudy days. The sunny-day-only caveat in the abstract is therefore incomplete and the parity claim is unsupported by the paper's own evidence.
[Section 5 (Tables 2, 3, 4, 5)] The paper reports three different overall RMSE values for the ViT: 2.61 in the pilot (Table 3), 3.38 as the 'best model' (Table 2), and 3.35 for the frozen-layer full-data model (Table 5), while Table 4 lists 'XX' placeholders for the sunny and cloudy categories despite reporting an overall value of 2.61. The manuscript never explains which result is the definitive one, leaving the reader unable to determine the model's actual performance. This internal inconsistency undermines the paper's main empirical claim.
[Section 3 vs Section 4.1; Table 1 provenance] The comparison between the ViT and the SUNSET baseline is not a controlled experiment. Section 3 states that the test set consists of 20 days (13,689 pairs, 10 sunny and 10 cloudy), but Section 4.1 describes a test set of 'ten days (five cloudy and five sunny)'. Moreover, the baseline RMSE values in Table 1 are quoted from prior work rather than recomputed in this pipeline: the paper notes only that 'This is the same test dataset used by the baseline SUNSET model.' The ViT pipeline resizes 256x256 images to 224x224 inside the HuggingFace feature extractor and uses a shuffled 95/5 train/validation split, whereas the baseline operated on 64x64 images under an unstated split and preprocessing regime. Without re-running the baseline under the same conditions, the reported RMSE gap could be a pipeline artifact rather than an architecture comparison, and the paper's parity conclusion rests on an unverified commensurability assumption.
[Section 5, pilot experiment (Table 3)] The pilot result of RMSE 2.61 is based on a single day of data (654 image-output pairs) with no hyperparameter tuning, and the paper describes the split as '90% − 5%', which is arithmetically incomplete. This pilot cannot support the inference in Section 5 that 'transformers are well-suited to our PV power output regression task,' nor can it substantiate the abstract's parity claim. The subsequent full-data runs produce different and worse overall RMSE values, so the pilot appears to be a small-sample artifact rather than a reliable measure of performance.

minor comments (7)

[Section 4.1] The description of the test set as 'ten days (five cloudy and five sunny)' directly contradicts Section 3's '20 days of image-PV output pairs... 10 of these days are classified as sunny and the other 10 as cloudy'; the paper should reconcile these counts and specify which test set was actually used for each reported result.
[Section 5.1] The sentence 'Using 654 samples... split 90% − 5% into training and validation sets' is incomplete: the two percentages sum to 95%, leaving 5% unattributed; in addition, the split description differs from the 95/5 split stated in Section 3.
[Table 4] Table 4 contains 'XX' placeholders for the sunny and cloudy RMSE values; a published table should either report the actual values or the table should be removed, since its presence signals that the experimental results are incomplete.
[Throughout, Figures 4, 5, and 6] Several figure references are garbled: Section 5.2 says 'The model performance is presented in 4' without a figure number, Section 5.3 refers to 'Table The model performance' and 'figure 5.3', and the captions for Figures 7, 8, and 9 are not mentioned in the text.
[Section 6 (Conclusion)] The Conclusion section is incomplete: it begins with 'Additionally, we may incorporate a CNN component...' and does not actually summarize the findings, restate the main conclusion, or discuss implications; the paper appears to be missing its concluding paragraphs.
[Title page and Section 5.2] There are typographical and grammatical errors, including 'Univeristy' twice in the author affiliations, 'great improves' in Section 5.2, and the special character rendering in reference [7] ('Szu˝cs'); these should be corrected in revision.
[Section 4] The phrase 'using a mean-squared error loss on a one-class classification' is not standard terminology for a regression task with a single continuous output; the paper should describe the model as a single-output regression head with MSE loss.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark against an external published baseline, not a derivation that reduces to its own inputs.

full rationale

The paper contains no derivation chain of the kind that could be circular. Its central claim is an experimental comparison: a frozen pretrained HuggingFace ViT (with ImageNet-pretrained weights) is applied to a PV-output regression task, and the resulting test RMSE values are compared with RMSE values for the SUNSET CNN baseline quoted from prior published work by the Stanford EAO lab (Sun, Szűcs, and Brandt; Sun, Venugopal, and Brandt). The SUNSET numbers are externally published results by different authors, not outputs of this paper's model or functions of its fitted parameters, so they function as an independent benchmark rather than a self-citation chain. The ViT weights are also external (ImageNet), so the 'prediction' is not defined in terms of the target quantity it is supposed to predict. No parameter is fitted to a subset and then renamed as a prediction, no uniqueness theorem is imported from same-author work, and no ansatz is smuggled in through a citation. Internal inconsistencies in the reported ViT RMSE values (Table 3: 2.61, Table 2: 3.38, Table 5: 3.35), the XX placeholders in Table 4, and the conflicting 10-day versus 20-day test-set descriptions are correctness and reproducibility concerns, not circularity. Therefore no significant circularity is present, and the score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper contains no derivation, so the ledger records the modeling choices and comparability assumptions that carry the RMSE comparison. Four hand-set choices (learning rate, batch size, frozen-layer regime, image resolution) plus four domain assumptions (image sufficiency, baseline commensurability, shuffled i.i.d., visual test classification) support the entire claim. No new entities are introduced. The heaviest burden is the assumption that the quoted SUNSET numbers were computed under exactly the same conditions as the ViT numbers, since the baseline is never re-run here.

free parameters (4)

Learning rate = 1e-5
Tuned over [1e-6, 5e-6, 1e-5, 5e-5, 1e-4] in Section 5.3; the headline Table 5 result uses 1e-5. Reported RMSE depends on this choice with no sensitivity analysis or seed repeats.
Batch size = 10
Tuned over {10, 64}; Table 5 reports batch size 10 as best. No evidence is given that the conclusion is robust to this choice.
Frozen-layer configuration = All layers frozen except final layer norm, pooler layer, and fully connected head
Adopted in Section 5.3 to cut training time, not selected by validation performance. The unfrozen warm-start run (Table 4) reports a different overall RMSE (2.61) than the frozen run (3.35), so the architecture-level claim depends on a training-speed compromise.
Image resolution = 256x256 training images, 64x64 test images, rescaled to 224x224 internally
Set in Section 3 without an ablation; the train/test resolution mismatch is unexplained and could contribute to the RMSE gap versus the baseline.

assumptions (4)

domain assumption Sky image content at snapshot time suffices to predict PV power output.
The Nowcast task in Sections 3 and 4 assumes a learnable map from fish-eye sky image to the paired power reading.
domain assumption The imported SUNSET baseline RMSE values were measured on the identical test set under identical preprocessing and normalization.
Section 3 asserts 'This is the same test dataset used by the baseline SUNSET model' and Table 1 imports 0.52/3.38/2.36 without re-running the baseline in the ViT pipeline.
domain assumption Shuffling the training pairs makes the i.i.d. assumption valid.
Section 3 shuffles before a 95/5 split specifically to suppress temporal patterns, but operational nowcasting is sequential; the shuffled split ignores autocorrelation and seasonality in cloud cover.
domain assumption Visual sunny/cloudy classification of the 20 test days is reliable.
Section 3 labels the test days 'by visual inspection'; the paper's central sunny-day finding (2.62 vs 0.52) depends on this subjective grouping.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Transformers Applied to Short-term Solar PV Power Output Forecasting." pith.science (2026). https://pith.science/paper/4BNJU3EH

@misc{pith2026250503188,
  author       = {Pith},
  title        = {Pith review of: Transformers Applied to Short-term Solar PV Power Output Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4BNJU3EH}},
  note         = {Machine review of arXiv:2505.03188}
}

read the original abstract

Reliable forecasts of the power output from variable renewable energy generators like solar photovoltaic systems are important to balancing load on real-time electricity markets and ensuring electricity supply reliability. However, solar PV power output is highly uncertain, with significant variations occurring over both longer (daily or seasonally) and shorter (within minutes) timescales due to weather conditions, especially cloud cover. This paper builds on existing work that uses convolutional neural networks in the computer vision task of predicting (in a Nowcast model) and forecasting (in a Forecast model) solar PV power output (Stanford EAO SUNSET Model). A pure transformer architecture followed by a fully-connected layer is applied to one year of image data with experiments run on various combinations of learning rate and batch size. We find that the transformer architecture performs almost as well as the baseline model in the PV output prediction task. However, it performs worse on sunny days.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 9 canonical work pages

[1]

Hypotheses for primary energy use, electricity use and co2 emissions of global computing and its shares of the total between 2020 and

Anders SG Andrae. Hypotheses for primary energy use, electricity use and co2 emissions of global computing and its shares of the total between 2020 and

work page 2020
[3]

Images are then down- sampled to a desired size (64 × 64, 128 × 128, 256 × 256, etc.)

Data The data set consists of one year (2017) of video record- ings of the sky above a PV solar array, pre-processed to snapshot minutely, high resolution images of daytime sky, each of size 2048 × 2048 pixels. Images are then down- sampled to a desired size (64 × 64, 128 × 128, 256 × 256, etc.). The sky images are collected in-house via a fish- eye camer...

work page 2017
[4]

This motivated us to as- sess the performance of the vision transformer to our solar power forecasting problem

Methods Within the past year, transformer architectures have gen- eralized from applications in the natural language process- ing field to computer vision. This motivated us to as- sess the performance of the vision transformer to our solar power forecasting problem. Initially, we considered a hy- brid CNN-transformer approach. However, Vaswani et al (2017)

work page 2017
[7]

The use of trans- formers for computer vision tasks is also quite recent [10]

and Forecast [8], [9]. The use of trans- formers for computer vision tasks is also quite recent [10]. Dosovitskiy et. al. (2020)

work page 2020
[8]

Convolutional neural net- works applied to sky images for short-term solar irradiance forecasting

Quentin Paletta and Joan Lasenby. Convolutional neural net- works applied to sky images for short-term solar irradiance forecasting. arXiv preprint arXiv:2005.11246,

arXiv 2005
[9]

Cloudy 30 20 10 0 0 1000 2000 3000 4000 5000 6000 Figure

work page 2000
[10]

A survey on visual transformer

Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chun- jing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv preprint arXiv:2012.12556,

arXiv 2012
[11]

An image is worth 16x16 words: Trans- formers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929,

arXiv 2010

Show all 13 references

[12]

Attention is all you need

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Il- lia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762,

arXiv
[13]

We therefore applied a pure transformer architecture to the solar forecasting problem

show that the transformer architecture is able to utilise the attention mechanism as an alternative to convolu- tions. We therefore applied a pure transformer architecture to the solar forecasting problem. The existing model from the EAO lab group (SUNSET) relies purely on CNN...

2020
[14]

for our implementation. This model has been pre-trained on the ImageNet-21k dataset (≈ 14 million images) at a resolution of 224 × 224 and finetuned on the ImageNet 2012 dataset (≈ 1 million images) at the same resolution. The model we used took in patches of resolution 16 × 1...

2012
[15]

Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau- mond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Re´mi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, M...

2020
[2019]

6 20 10 0 0 1000 2000 3000 4000 5000 6000 7000 Figure

2000

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Hypotheses for primary energy use, electricity use and co2 emissions of global computing and its shares of the total between 2020 and

Anders SG Andrae. Hypotheses for primary energy use, electricity use and co2 emissions of global computing and its shares of the total between 2020 and

work page 2020

[2] [3]

Images are then down- sampled to a desired size (64 × 64, 128 × 128, 256 × 256, etc.)

Data The data set consists of one year (2017) of video record- ings of the sky above a PV solar array, pre-processed to snapshot minutely, high resolution images of daytime sky, each of size 2048 × 2048 pixels. Images are then down- sampled to a desired size (64 × 64, 128 × 128, 256 × 256, etc.). The sky images are collected in-house via a fish- eye camer...

work page 2017

[3] [4]

This motivated us to as- sess the performance of the vision transformer to our solar power forecasting problem

Methods Within the past year, transformer architectures have gen- eralized from applications in the natural language process- ing field to computer vision. This motivated us to as- sess the performance of the vision transformer to our solar power forecasting problem. Initially, we considered a hy- brid CNN-transformer approach. However, Vaswani et al (2017)

work page 2017

[4] [7]

The use of trans- formers for computer vision tasks is also quite recent [10]

and Forecast [8], [9]. The use of trans- formers for computer vision tasks is also quite recent [10]. Dosovitskiy et. al. (2020)

work page 2020

[5] [8]

Convolutional neural net- works applied to sky images for short-term solar irradiance forecasting

Quentin Paletta and Joan Lasenby. Convolutional neural net- works applied to sky images for short-term solar irradiance forecasting. arXiv preprint arXiv:2005.11246,

arXiv 2005

[6] [9]

Cloudy 30 20 10 0 0 1000 2000 3000 4000 5000 6000 Figure

work page 2000

[7] [10]

A survey on visual transformer

Kai Han, Yunhe Wang, Hanting Chen, Xinghao Chen, Jianyuan Guo, Zhenhua Liu, Yehui Tang, An Xiao, Chun- jing Xu, Yixing Xu, et al. A survey on visual transformer. arXiv preprint arXiv:2012.12556,

arXiv 2012

[8] [11]

An image is worth 16x16 words: Trans- formers for image recognition at scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint arXiv:2010.11929,

arXiv 2010