Pith. sign in

REVIEW 3 major objections 6 minor 29 references

Meteosat Third Generation imagery improves CNN-based SSI retrieval

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Higher-resolution MTG/FCI imagery improves CNN surface solar irradiance retrieval under cloudy skies, but not under clear skies.

desk verdict Solid first look at MTG/FCI for CNN SSI retrieval: real cloudy-sky gains vs MSG-only and SARAH-3, honest clear-sky null, but the hybrid edge is partly confounded by same-year fine-tune/eval. read the letter →

arxiv 2607.28093 v1 pith:X43QLPD3 submitted 2026-07-30 physics.ao-ph cs.LGphysics.data-an

Gordei Pribõtkin , Piia Post , Velle Toll This is my paper
classification physics.ao-phcs.LGphysics.data-an
keywords surfacesolarirradianceMTG/FCIMSG/SEVIRIconvolutionalneuralnetworkmulti-imagerfusionSARAH-3clear-skyindexphotovoltaicmonitoring
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 the newer, higher-resolution Meteosat Third Generation FCI imager actually helps machine-learning models estimate how much sunlight hits the ground, compared with relying only on older Meteosat Second Generation SEVIRI data. The authors build a multi-branch convolutional network that can fuse both imagers at their native resolutions, train it on Estonian pyranometer stations with site-based cross-validation, and compare against the physics-based SARAH-3 product. They find a clear win when clouds dominate: the hybrid model cuts RMSE by about 8 W/m² in overcast and 6 W/m² in cloudy conditions and scores roughly 20–35% skill versus SARAH-3. Under partly cloudy or clear skies the extra resolution adds no significant gain, and both neural models still lose to SARAH-3 in clear air. The practical message is that next-generation geostationary imagery is worth fusing for cloudy regimes that matter most to PV variability, yet spatial resolution alone does not fix the long-standing clear-sky weakness of learning-based retrieval.

What carries the argument

The multi-branch multi-resolution CNN: separate ResNet-style branches encode SEVIRI 3 km/1 km and FCI 1 km/0.5 km patches, embeddings are concatenated with solar-geometry and clear-sky features plus missingness flags, and SEVIRI branches are frozen while FCI branches and the head are fine-tuned on the short FCI year.

What would settle it

Retrain and re-evaluate the hybrid versus SEVIRI-only models once a second full independent FCI year exists, holding out that year entirely from fine-tuning; if the cloudy-sky RMSE gap and skill scores shrink or vanish, the claimed FCI benefit was inflated by same-year leakage.

Watch

Extended reading notes

Core claim

A hybrid multi-imager, multi-resolution CNN that freezes long-trained SEVIRI branches and fine-tunes FCI branches significantly outperforms a SEVIRI-only CNN under overcast and cloudy conditions (RMSE reductions of 8.2 and 5.7 W m−2) and beats SARAH-3 with skill scores of 35%, 21%, and 20% in overcast, cloudy, and all-sky conditions, while showing no significant RMSE gain under partly cloudy or clear skies and underperforming SARAH-3 in clear skies.

Load-bearing premise

That training and testing the hybrid model on the same single year of FCI data, with full temporal overlap and shared weather patterns across sites, still gives an unbiased measure of the true benefit of the newer imager.

Editorial extensions

If this is right

  • Fusing abundant older SEVIRI archives with limited new FCI data is a workable path to better cloudy-sky SSI maps for PV monitoring in regions like Northern Europe.
  • Higher spatial resolution alone will not close the clear-sky gap versus physics products; bias and seed instability under clear skies remain the binding limits.
  • Site-based cross-validation with multiple seeds is needed to claim imager gains; single-model single-split results can misstate cloudy versus clear performance.
  • The same multi-branch embeddings are a natural base for short-lead SSI forecasting by adding temporal modules over the branch features.

Reading between the lines

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

  • Clear-sky failure may be driven more by rare clear samples (only ~4% of the set) and resulting calibration bias than by missing aerosol bands alone, which would explain the high seed-to-seed RMSE scatter only on clear days.
  • If the cloudy-sky gain holds on independent years, hybrid ML products could become the preferred operational source precisely where PV ramp risk is highest.
  • Extending the architecture beyond Estonia will test whether the FCI benefit survives larger albedo and aerosol diversity that pure geographic splits expose.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents a multi-imager, multi-resolution CNN for 10-minute SSI retrieval over Estonia, combining MSG/SEVIRI (1–3 km) and MTG/FCI (0.5–1 km) imagery with solar-geometry and clear-sky features. Using 8-fold site-held-out cross-validation with five seeds per fold, the authors compare a SEVIRI-only model (trained 2021–2025) to a HYBRID model that freezes SEVIRI branches and fine-tunes FCI branches plus the head on 2025 FCI data. Against pyranometer targets and the SARAH-3 baseline (matched timestamps), HYBRID significantly reduces RMSE versus SEVIRI-only under overcast (−8.2 W m−2) and cloudy (−5.7 W m−2) skies and yields skill scores of 35%, 21%, and 20% versus SARAH-3 in overcast, cloudy, and overall conditions; no significant gain appears under partly cloudy or clear skies, and both ML models underperform SARAH-3 in clear skies. The authors conclude that higher-resolution FCI helps when clouds dominate irradiance variability, but resolution alone does not fix clear-sky ML limitations.

Significance. If the cloudy-sky FCI benefit holds under cleaner controls, this is a timely and practically useful result: MTG/FCI has only recently become available, prior ML SSI work used coarser geostationary imagery, and a demonstrated hybrid strategy (long SEVIRI record + limited FCI fine-tuning) would matter for operational PV monitoring in Europe. Strengths include external pyranometer targets, an independent physics baseline (SARAH-3), sky-condition stratification, site CV with multiple seeds, fold-level SEM, Bonferroni-corrected paired tests, and per-fold appendix tables—more rigorous uncertainty treatment than much of the related ML-SSI literature. The clear-sky negative result is also informative and consistent with prior work. The main significance risk is that the reported FCI gain may partly reflect same-year fine-tune/eval alignment and extra capacity rather than spatial resolution per se.

major comments (3)
  1. [Model §3; Table 5; Discussion] Model §3 and Fig. 2b / Data §2: The central HYBRID vs SEVIRI-only contrast (Table 5: −8.2 and −5.7 W m−2 under overcast/cloudy) is not a clean imager or resolution contrast. SEVIRI-only is trained on 2021–2025; HYBRID freezes those branches and updates only FCI branches + head on calendar year 2025—the same year that defines all test samples. Shared weather-pattern leakage and possible 2025-specific recalibration of the head are acknowledged in Discussion but not controlled. A load-bearing control is needed: e.g., fine-tune only the SEVIRI-only head (and/or a SEVIRI branch) on 2025 with no FCI input, and/or train a matched-capacity SEVIRI model on the 2025 window, then re-run the paired tests. Without this, the claim that MTG/FCI imagery (rather than same-year adaptation or extra parameters) drives the cloudy-sky gain remains under-supported.
  2. [Abstract; Table 1; §3; Discussion] Abstract, Introduction, and Discussion attribute the HYBRID gain primarily to “higher spatial resolution,” but Table 1 and §3 show non-identical spectral sets, different patch physical sizes (51 km vs 20.5 km for 0.5 km bands), oversampling at oblique view, and additional FCI branch capacity. Discussion correctly lists these confounds yet the title and headline claims still single out resolution. Either add ablations that hold spectrum/capacity closer to fixed (e.g., FCI bands downsampled to SEVIRI resolution; SEVIRI-only with extra dummy branches) or revise claims to “adding MTG/FCI via hybrid fine-tuning improves cloudy-sky retrieval,” without asserting resolution as the isolated cause.
  3. [§4 Evaluation; Table 3; Tables 5–7] Evaluation §4 and Table 3: Sky classes are defined by manually tuned CSI_day and VI thresholds (clear = only 4.4% of samples). Results and skill scores are stratified on these labels, and clear-sky instability/bias conclusions depend on them. Sensitivity of Table 5–7 rankings to modest threshold changes (and to an alternative classifier, e.g., instantaneous CSI or cloud mask) should be reported so the overcast/cloudy significance claims are not brittle to the ad-hoc cutoffs.
minor comments (6)
  1. [§3 Model, Eqs. (4)–(5)] Eq. (4)–(5): CSI* uses ε = 10 W m−2 and SSI_cs normalization by 800 W m−2; briefly justify these scales (or show insensitivity) so they are not free knobs affecting RMSE in low-sun regimes.
  2. [§3 Model] SEVIRI 3 km patches are limited to 17 px “to mitigate data leakage” from overlapping sites (§3); state typical inter-station distances and residual overlap risk for 51 px / 1 km and FCI patches.
  3. [Table 6; §5 Results] Table 6 skill in clear skies has very large SEM (±23–26%); the text already notes non-significance—consider emphasizing that the −38% figure is unstable rather than a precise deficit.
  4. [Figures 1 and 5] Figure 1 is conceptual and fine; ensure axis labels and condition names in Fig. 5 match Table 5 units (W m−2) consistently in the final layout.
  5. [Front matter; Discussion] Typos/consistency: “V elle Toll” spacing in author list; “raining” → “training” in Discussion (“importance of clear-sky samples could be increased during raining”); arXiv date stamp “30 Jul 2026” looks like a placeholder.
  6. [§1 Introduction] Related work is appropriate; a brief explicit comparison of temporal resolution (10 min here vs hourly in some prior NN studies) in the introduction would help readers place the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: external pyranometer targets, independent SARAH-3 baseline, and standard CSI*/skill metrics; confounds are validity issues, not definitional loops.

full rationale

This is an empirical CNN retrieval paper. The load-bearing claims are RMSE differences (HYBRID vs SEVIRI-only; both vs SARAH-3) under sky-condition strata, obtained by training on satellite patches plus solar-geometry/clear-sky features and evaluating against held-out station pyranometer SSI. The target CSI* = SSI/(SSI_cs+ε) is a standard stabilized clear-sky index used only as a training transform; SSI is recovered by inverting it and compared to external measurements. Skill score SS = 1 − RMSE_model/RMSE_SARAH-3 is a conventional relative-error metric against an independent Heliosat product. Site-based CV, multiple seeds, and Bonferroni paired tests do not force the reported cloudy-sky gap by construction. Same-year FCI fine-tune/eval overlap, shared weather-pattern leakage, extra HYBRID parameters, and manually tuned VI/CSI_day bins are experimental-validity and attribution concerns, not self-definitional or fitted-input-as-prediction circularity. No self-citation uniqueness theorem or renamed known identity underpins the central claim. Derivation chain is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

Empirical ML study: claim rests on standard remote-sensing and ML practice plus several paper-specific design choices (patch geometry, sky taxonomy, hybrid freeze protocol, clear-sky model). No new physical entities. Free parameters are training hyperparameters and manually set classification cutoffs that define the stratified results the abstract highlights.

free parameters (6)
  • Sky-condition CSI_day and VI thresholds = e.g. clear: CSI_day>0.7 and 0.8<VI<1.2; overcast: CSI_day<0.5 (plus exclusions)
    Clear/partly/cloudy/overcast bins defined by manually reviewed CSI_day and VI limits (Table 3); these bins carry the headline stratified RMSE and skill claims.
  • Stabilization epsilon in CSI* = 10 W m−2
    ε=10 W m−2 in CSI*=SSI/(SSI_cs+ε); changes target scale near low sun.
  • SSI_cs normalization scale = 800 W m−2
    Clear-sky feature scaled by 800 W m−2.
  • Patch physical sizes and pixel grids = 51×51 px (most), 17×17 px (SEVIRI 3 km); 51/20.5 km
    51 km / 20.5 km patches; 17 px for SEVIRI 3 km to limit site overlap leakage—design choices that set receptive field and leakage control.
  • Gaussian coordinate encoding scale = 0.25 in denominator
    E_gauss=exp(-(x²+y²)/0.25) width chosen by authors.
  • Training hyperparameters (LR, weight decay, epochs, patience) = LR 1e-3 / 5e-5; wd 1e-4; patience 5
    SEVIRI LR 1e-3, HYBRID fine-tune 5e-5, AdamW wd 1e-4, max 35 epochs, early stop patience 5 on rolling val RMSE—affect which checkpoint is reported.
assumptions (6)
  • domain assumption Ineichen/Perez clear-sky model (pvlib) is an adequate SSI_cs reference for CSI* targets and features.
    Used for targets, features, and VI/CSI_day sky typing throughout §§2–4.
  • ad hoc to paper Site-held-out CV with shared 2025 calendar across train/test still supports generalization claims for FCI benefit.
    Authors acknowledge weather-pattern leakage (§2) but still report test metrics as primary evidence.
  • ad hoc to paper Freezing SEVIRI branches and training only FCI branches+head on one year isolates complementary FCI information without unfairly favoring HYBRID.
    Training protocol Table 2 / §3; Discussion notes possible same-year bias.
  • domain assumption Pyranometer 10-minute SSI at eight Estonian sites is ground truth for pixel-centered satellite patches.
    Standard point-vs-satellite mismatch accepted as evaluation basis (§2, §4).
  • domain assumption ResNet-style multi-branch CNNs plus solar geometry features are appropriate function classes for SSI retrieval.
    Architecture justified by prior SSI CNN literature (§3).
  • standard math Bonferroni-corrected paired t-tests on fold-level mean RMSEs adequately control multiplicity for claimed significance.
    §5 Tables 5–6 statistical protocol.
invented entities (2)
  • HYBRID SEVIRI–FCI multi-resolution multi-branch CNN with modality missingness flags
    purpose: Fuse long SEVIRI record with scarce high-res FCI for 10-minute SSI*
    Model class is engineered for this study; not a new physical object. Independent evidence is empirical hold-out RMSE only.
  • Stabilized clear-sky index CSI* with ε=10 W m−2
    purpose: Regression target to stabilize low-sun ratios
    Minor variant of standard CSI; ε chosen in-paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Meteosat Third Generation imagery improves CNN-based SSI retrieval." pith.science (2026). https://pith.science/paper/X43QLPD3

@misc{pith2026260728093,
  author       = {Pith},
  title        = {Pith review of: Meteosat Third Generation imagery improves CNN-based SSI retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X43QLPD3}},
  note         = {Machine review of arXiv:2607.28093}
}
abstract

Accurate Surface Solar Irradiance (SSI) estimation is increasingly important for photovoltaic energy monitoring and forecasting. The recently introduced Meteosat Third Generation (MTG) satellite constellation provides imaging data with higher spatial resolution compared to the Meteosat Second Generation (MSG) satellite constellation, but its benefits for machine-learning-based SSI retrieval have not been well established. In this work, we introduce a multi-imager and multi-resolution convolutional neural network architecture for 10-minute SSI retrieval over Northern Europe (Estonia) using MSG/SEVIRI and MTG/FCI satellite imagery together with solar-geometry and clear-sky irradiance features. Model performance is evaluated against ground-based pyranometer measurements from eight Estonian meteorological stations using site-based cross-validation and multiple training seeds. Model performance is also compared with the SARAH-3 physics-based satellite SSI product. The hybrid SEVIRI-FCI model significantly outperformed the SEVIRI-only model under overcast and cloudy conditions, reducing RMSE by 8.2 W m$^{-2}$ and 5.7 W m$^{-2}$, respectively. However, under partly cloudy or clear skies, no statistically significant difference in RMSE was observed between the SEVIRI-FCI hybrid and the SEVIRI-only models. Compared with physics-based SARAH-3, the hybrid model yielded skill scores of 35 % under overcast conditions, 21 % under cloudy conditions, and 20 % overall. Furthermore, both models underperformed SARAH-3 in clear-sky conditions. These results show that higher-resolution MTG/FCI imagery improves CNN-based SSI retrieval when clouds dominate irradiance variability, but also indicate that higher spatial resolution alone is insufficient to address clear-sky limitations in machine-learning-based SSI retrieval.

Figures

Figures reproduced from arXiv: 2607.28093 by the authors.

Figure 1
Figure 1. Conceptual study design. We use the combination of MSG and MTG data. Our dataset spans 5 years for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Diagrams showing spatial and temporal extent of the dataset. (a) Map showing locations of meteorological [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Diagram of the ML-based SSI retrieval model. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Diagram of the Branch CNN module. In this diagram, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Paired RMSE difference (HYBRID – SEVIRI-only) by sky condition and overall. Lower value indicates [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Example clear-sky and cloudy-sky samples for Tallinn. Models were taken from one of the seeds and the fold [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 1 linked inside Pith

  1. [1]

    Trends in photovoltaic applications

    Gaëtan Masson, Melodie de l’Epine, Izumi Kaizuka, and Junko Okazaki. Trends in photovoltaic applications

  2. [2]

    Manajit Sengupta, Aron Habte, Stefan Wilbert, Christian Gueymard, Jan Remund, Elke Lorenz, Wilfried van Sark, and Adam R. Jensen. Best practices handbook for the collection and use of solar resource data for solar energy applications: Fourth edition. Technical report, IEA Photovoltaic Power Systems Programme, 2024

    2024

  3. [3]

    Performance assessment and dynamic fault detection in photovoltaic systems using artificial intelligence.Energy, 330:136759, 2025

    Nelson Salazar-Peña, Alejandra Tabares, and Andrés González-Mancera. Performance assessment and dynamic fault detection in photovoltaic systems using artificial intelligence.Energy, 330:136759, 2025

    2025

  4. [4]

    The method heliosat-2 for deriving shortwave solar radiation from satellite images.Solar Energy, 77(2):159–169, 2004

    C Rigollier, M Lefèvre, and L Wald. The method heliosat-2 for deriving shortwave solar radiation from satellite images.Solar Energy, 77(2):159–169, 2004

    2004

  5. [5]

    Pfeifroth, J

    U. Pfeifroth, J. Drücke, S. Kothe, J. Trentmann, M. Schröder, and R. Hollmann. Sarah-3 – satellite-based climate data records of surface solar radiation.Earth System Science Data, 16(11):5243–5265, 2024

    2024

  6. [6]

    Cams gridded solar radiation, 2022

    Copernicus Atmosphere Monitoring Service (CAMS). Cams gridded solar radiation, 2022. Accessed 2026-02-17

    2022

  7. [7]

    User guide to the cams radiation service (crs): Status december 2021

    Marion Schroedter-Homscheidt. User guide to the cams radiation service (crs): Status december 2021. User guide / technical report CAMS2_73_2021SC1_D3.2.1_2021_UserGuide_v1, Copernicus Atmosphere Monitoring Service (CAMS) / Deutsches Zentrum für Luft- und Raumfahrt (DLR), December 2021. Issued by DLR; dated 2021-12-19. Accessed 2026-02-17

    2021

  8. [8]

    The national solar radiation data base (nsrdb).Renewable and Sustainable Energy Reviews, 89:51–60, June 2018

    Manajit Sengupta, Yu Xie, Anthony Lopez, Aron Habte, Galen Maclaurin, and James Shelby. The national solar radiation data base (nsrdb).Renewable and Sustainable Energy Reviews, 89:51–60, June 2018

    2018

Show all 29 references
  1. [9]

    B. K. Bhattacharya and N. Rahul.Surface Insolation over Land from INSAT-3D, pages 335–360. MOSDAC / ISRO, May 2015. Algorithm Theoretical Basis Document (ATBD)

    2015

  2. [10]

    European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), v8 e-signed edition

    EUMETSAT.MSG Level 1.5 Image Data F ormat Description. European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT), v8 e-signed edition

  3. [11]

    Published 29 Jan 2026

    EUMETSAT.MTG FCI Level 1c Data Guide, version 1.8 edition, January 2026. Published 29 Jan 2026. Accessed 2026-06-29

    2026

  4. [12]

    Verbois, Y .-M

    H. Verbois, Y .-M. Saint-Drenan, V . Becquet, B. Gschwind, and P. Blanc. Retrieval of surface solar irradiance from satellite imagery using machine learning: pitfalls and perspectives.Atmospheric Measurement Techniques, 16(18):4165–4181, 2023

    2023

  5. [13]

    K. R. Schuurman and Angela Meyer. Surface solar radiation: Ai satellite retrieval can outperform heliosat and generalizes to other climate zones.International Journal of Remote Sensing, 46(8):3331–3362, 2025

    2025

  6. [14]

    EUMETSAT releases pre-operational data from MTG imager, 2024

    EUMETSAT. EUMETSAT releases pre-operational data from MTG imager, 2024. Published 23 September 2024; last updated 24 September 2024

    2024

  7. [15]

    Deep learning and process understanding for data-driven earth system science.Nature, 566(7743):195– 204, 2019

    Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, and Prabhat. Deep learning and process understanding for data-driven earth system science.Nature, 566(7743):195– 204, 2019. 13

    2019

  8. [16]

    A deep learning algorithm to estimate hourly global solar radiation from geostationary satellite data.Renewable and Sustainable Energy Reviews, 114:109327, 2019

    Hou Jiang, Ning Lu, Jun Qin, Wenjun Tang, and Ling Yao. A deep learning algorithm to estimate hourly global solar radiation from geostationary satellite data.Renewable and Sustainable Energy Reviews, 114:109327, 2019

    2019

  9. [17]

    Global and direct solar irradiance estimation using deep learning and selected spectral satellite images.Applied Energy, 352:121979, 2023

    Shanlin Chen, Chengxi Li, Yuying Xie, and Mengying Li. Global and direct solar irradiance estimation using deep learning and selected spectral satellite images.Applied Energy, 352:121979, 2023

    2023

  10. [18]

    Rapid scan high rate seviri level 1.5 image data - msg (msg15-rss), 2009

    EUMETSAT. Rapid scan high rate seviri level 1.5 image data - msg (msg15-rss), 2009. Collection ID: EO:EUM:DAT:MSG:MSG15-RSS. Accessed 2026-02-13

    2009

  11. [19]

    Fci level 1c high resolution image data - mtg - 0 degree, 2024

    EUMETSAT. Fci level 1c high resolution image data - mtg - 0 degree, 2024. Product ID: EO:EUM:DAT:0665. Accessed 2026-02-13

    2024

  12. [20]

    Surface radiation data set - heliosat (sarah) - edition 3

    EUMETSAT. Surface radiation data set - heliosat (sarah) - edition 3. Collection ID: EO:EUM:DAT:0863. Accessed 2026-02-13

    2026

  13. [21]

    Holmgren, Clifford W

    William F. Holmgren, Clifford W. Hansen, and Mark A. Mikofski. pvlib python: a python package for modeling solar energy systems.Journal of Open Source Software, 3(29):884, 2018

    2018

  14. [22]

    A new airmass independent formulation for the linke turbidity coefficient

    Pierre Ineichen and Richard Perez. A new airmass independent formulation for the linke turbidity coefficient. Solar Energy, 73(3):151–157, 2002

    2002

  15. [23]

    EPSG:3301 – Estonian Coordinate System of 1997

    EPSG Geodetic Parameter Dataset. EPSG:3301 – Estonian Coordinate System of 1997

    1997

  16. [24]

    Multimodal deep learning

    Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and Andrew Ng. Multimodal deep learning. pages 689–696, 01 2011

    2011

  17. [25]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015

    2015

  18. [26]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis Bach and David Blei, editors,Proceedings of the 32nd International Conference on Machine Learning, volume 37 ofProceedings of Machine Le...

    2015

  19. [27]

    Mish: A self regularized non-monotonic neural activation function.CoRR, abs/1908.08681, 2019

    Diganta Misra. Mish: A self regularized non-monotonic neural activation function.CoRR, abs/1908.08681, 2019

    1908 arXiv

  20. [28]

    Stein, Clifford W

    Joshua S. Stein, Clifford W. Hansen, and Matthew J. Reno. The variability index: A new and novel metric for quantifying irradiance and pv output variability. 2012. 14 A Performance metrics by fold Table 8: Per-fold RMSE for each held-out test site, by sky condition, for the SE...

    2012

  21. [2025]

    Technical report, IEA Photovoltaic Power Systems Programme, 2025

    2025

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.