acceptodds
Under review as a conference paper at ICLR 2027

PVFM: A Solar-Aware, Weather-Conditioned Foundation Model for Zero-Shot Photovoltaic Forecasting

Abstract

Rapid PV deployment across regions and installation types demands accurate and scalable power forecasting, particularly for newly commissioned plants. Existing forecasters are often site- and task-specific and transfer poorly to unseen sites, while general-purpose TSFMs show limited advantages in short-horizon and high-resolution PV forecasting, leaving unified forecasting across diverse PV scenarios an open challenge. We introduce PVFM, a 3.88M-parameter weather-conditioned foundation model for zero-shot PV power forecasting. The model separately models historical observations and future weather covariates through dedicated information streams, augmented with solar-aware embeddings that encode solar geometry and static identity information. It adopts a general-to-specific strategy, first pretraining on 59.2 million PV observations across global sites with weather covariates from multiple sources. The pretrained model is further evaluated under ramp events and extended through multimodal fine-tuning with satellite imagery for short-horizon forecasting. Across five zero-shot tasks spanning short- to long-term horizons at 1-hour and 15-minute resolutions, PVFM consistently outperforms both full-shot models and existing TSFMs, reducing MAE by 10.2-24.3% and CRPS by 10.0-29.5%. Satellite-image fine-tuning further reduces RMSE by 9.1% in ultra-short-term forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Photovoltaic Power Forecasting, foundation models, time series analysis, probabilistic forecasting, zero-shot learning.

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3VemOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PVFM is a 3.88M-parameter model for zero-shot PV forecasting, pretrained on 59.2 million valid PV time points from 746 sites. It uses a historical encoder and a future NWP-conditioned decoder, with solar-aware temporal embeddings and site conditioning, to produce non-crossing quantile forecasts. Five tasks cover hourly and 15-minute data, with forecast horizons from four hours to seven days. The evaluation includes unseen sites, region holdouts in Hong Kong and Australia, and ramp events. A satellite-image adapter is also fine-tuned with the PVFM backbone frozen.

Critical strengths

I find the comparison with target-site full-shot models in Table 1 particularly useful. PVFM achieves lower MAE and CRPS without updating its parameters on the target sites. The Hong Kong and Australia holdouts provide a further test of transfer to regions excluded from pretraining. The authors also acknowledge that target-site training remains stronger on some Australian 15-minute tasks.

Critical weaknesses
  1. The satellite result does not separate the benefit of imagery from target-site adaptation or the change in training objective. Table 11 compares zero-shot PVFM with a model whose satellite-image adapter is fine-tuned on target-site labels using the masked mean squared error of the median forecast (Equation 38). An adapter trained in the same way without imagery is needed as a control. This concern is limited to the satellite extension. The reported results and model configurations also need clarification: Figure 6 and Table 11 give different RMSE values (0.0839 versus 0.0842) and R² values (0.8642 versus 0.8628). Rounding does not explain this difference. Table 5 suggests that PVFM-Sat uses a larger backbone than the main 3.88M model, but Table 11 does not identify the pretrained checkpoint for each PVFM row.

  2. I am also concerned about whether the weather inputs are available when the PV forecast is issued. Appendix A.2 aligns archived GFS forecasts by valid timestamp, but does not identify the forecast run selected at each origin. Historical weather inputs come from ERA5 reanalysis. Section 6.4 acknowledges its availability delay, but tests removal of ERA5 only for the Anhui satellite task. The comparisons using shared weather inputs remain informative. Without an explicit GFS selection rule and tests using historical weather available at forecast issuance, however, I cannot judge whether the reported zero-shot gains would hold in operational PV forecasting.

Questions for authors
  1. Can you compare the satellite-image adapter with an adapter without imagery, using the same pretrained checkpoint, target-site labels, training objective and training budget? Please also clarify the results and checkpoint configurations noted above.

  2. Which GFS forecast runs and recent historical weather inputs are available at each forecast origin after publication delays are considered? With those inputs, do the reported zero-shot gains over pretrained TSFMs and target-site full-shot models remain across regions, forecast horizons and both temporal resolutions?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 3Vem 1/3

W1 & Q1: Does satellite imagery provide additional benefits beyond target-site adaptation?

Weakness 1. “The satellite result does not separate the benefit of imagery from target-site adaptation or the change in training objective. ... This concern is limited to the satellite extension.”

Question 1. “... compare the satellite-image adapter with an adapter without imagery, using the same pretrained checkpoint, target-site labels, training objective and training budget?”

Response. We thank the reviewer for this suggestion. To distinguish the contribution of satellite imagery from target-site adaptation, we conducted additional experiments comparing fine-tuning with and without satellite imagery. We used GFS forecasts selected according to their actual initialization times, ensuring that only meteorological forecasts available at the prediction origin were used. The results are presented below.

Table 1. Contribution of Satellite Imagery to PV Forecasting Performance. Both adaptation settings use the same pretrained checkpoint, target-site training labels, loss function, training budget, and evaluation windows.

SettingNWPSatelliteMAE ↓RMSE ↓R² ↑
Zero-shot✓–0.04080.08790.8374
Fine-tuned✓–0.03910.08380.8521
Fine-tuned✓✓0.03400.07570.8793

Compared with PV+GFS fine-tuning without satellite imagery, satellite-image adaptation reduces MAE by 13.10% and RMSE by 9.67%, while improving R² by 3.20%.

These results demonstrate that satellite imagery can complement historical meteorological information by providing additional observations of local cloud conditions, thereby further improving short-term PV forecasting performance.


W2: Clarification of satellite results and pretrained checkpoint configurations

Weakness 2. “The reported results and model configurations also need clarification: Figure 6 and Table 11 give different RMSE values (0.0839 versus 0.0842) and R² values (0.8642 versus 0.8628). Rounding does not explain this difference. Table 5 suggests that PVFM-Sat uses a larger backbone than the main 3.88M model, but Table 11 does not identify the pretrained checkpoint for each PVFM row.”

Question 1 (configuration clarification). “Please also clarify the results and checkpoint configurations noted above.”

Response. We thank the reviewer for identifying these inconsistencies and requesting clarification of the model configurations. We have addressed these issues as follows.

1. Satellite results. We have updated the satellite fine-tuning results table to report the final results for fine-tuned PVFM: MAE = 0.0390, RMSE = 0.0839, and R² = 0.8642. These values are consistent with Fig. 6(c–d) and the 9.1% RMSE reduction reported in the abstract and Section 6.4.

2. Backbone configuration. We clarify that all PVFM variants in the satellite experiments use the same 3.88M-parameter backbone as the main experiments. We have corrected Table 5 to reflect the actual model configuration, with 4.503M total parameters including the satellite adapter.

3. Pretrained checkpoint configurations. We have revised the description and caption of the satellite fine-tuning results table to clarify the three PVFM configurations:

  • PVFM†: Zero-shot PVFM with historical ERA5 and future NWP.
  • PVFM‡: Zero-shot PVFM without historical ERA5, using future NWP.
  • Fine-tuned PVFM: PVFM adapted with historical satellite imagery and future NWP, without ERA5.

Revision. These changes have been incorporated into the revised manuscript, with the corrected parameter counts in Table 5 and further clarifications provided in Appendix C.4.2.

AuthorsOpenReview ↗

Response to Reviewer 3Vem 2/3

W3 & Q2: Are the weather inputs available at forecast issuance, and do PVFM’s gains persist under these constraints?

Weakness 3. “Appendix A.2 ... does not identify the forecast run selected at each origin. ... tests removal of ERA5 only for the Anhui satellite task. ... I cannot judge whether the reported zero-shot gains would hold in operational PV forecasting.”

Question 2. “Which GFS forecast runs and recent historical weather inputs are available at each forecast origin after publication delays are considered? ... do the reported zero-shot gains over pretrained TSFMs and target-site full-shot models remain across regions, forecast horizons and both temporal resolutions?”

Response. We thank the reviewer for raising concerns about the availability of historical and future meteorological inputs at forecast issuance. To address these concerns, we conducted additional evaluations incorporating GFS initialization times, publication delays, and historical weather availability, examining whether PVFM retains its zero-shot advantages under more realistic forecasting conditions.

1. Meteorological Input Availability at Forecast Issuance

(1) Future GFS inputs.

  • Original: For GDEX data before March 23, 2021, we used the daily 00 UTC GFS run with eight forecast lead times (+3, +6, ..., +24 h), linearly interpolated to hourly resolution. Subsequent data were obtained from Open-Meteo as continuous hourly GFS sequences. Both were aligned by valid time without explicitly enforcing forecast availability at each PV forecast origin.
  • Additional evaluation: We additionally retrieved complete GFS forecast cycles and selected the latest available cycle satisfying (r+6 h≤t)(r+6\,\mathrm{h}\leq t) among the downloaded forecasts, assuming a 6-hour publication delay. Each prediction window used a single forecast cycle without stitching.

(2) Historical weather inputs.

  • Original: We used ERA5 as a unified historical weather source for cross-site pretraining, as continuous ground-based observations were difficult to obtain across globally distributed PV sites. To address its publication delay, we also explored satellite observations for ultra-short-term forecasting.
  • Additional evaluation: Following the reviewer's concern, we further examined historical ERA5 availability by masking the most recent 6, 24, or 120 hours, or removing ERA5 entirely, and evaluated PVFM variants pretrained to handle missing historical weather inputs.

Due to time and download constraints, we selected a fixed subset of available GFS forecast cycles, covering 99 unseen sites and 14,214 windows across five tasks. All subsequent experiments used the same test windows and single-cycle GFS inputs.

2. Zero-Shot Performance Under Single-Cycle GFS Inputs

Finding-1. PVFM maintains its zero-shot advantage under single-cycle GFS inputs, even compared with full-shot models using the original GFS inputs.

We compared PVFM with iTransformer-TC (Full-shot) and CITRAS-FM (TSFM) under single-cycle GFS inputs. Despite using more realistic forecast-origin-constrained meteorological inputs, PVFM achieves lower MAE across all five tasks than both reference models. Notably, its five-task average MAE is 17.5% lower than that of iTransformer-TC evaluated with the original, valid-time-aligned GFS inputs, which provide more favorable meteorological information.

Table 2. Forecasting Performance under Original and Single-Cycle GFS Inputs. Original denotes valid-time-aligned GFS with complete historical ERA5; Single-cycle denotes forecast-origin-constrained GFS with historical ERA5 unchanged.

ModelFuture GFS1h 6h1h 24h1h 168h15min 4h15min 6hAvg. MAE ↓Avg. CRPS ↓
iTransformer-TC (Full-shot)Original0.05750.06380.08000.05190.05110.06090.0535
iTransformer-TC (Full-shot)Single-cycle0.05750.06520.08300.05670.05330.06320.0554
CITRAS-FM (TSFM)Single-cycle0.06030.07940.06850.08580.08000.07480.0566
PVFMSingle-cycle0.04950.05300.05630.04630.04600.05020.0363
AuthorsOpenReview ↗

Response to Reviewer 3Vem 3/3

W3 & Q2 (continued)

3. Zero-Shot Performance Under Historical Weather Availability Constraints

Finding-2. PVFM maintains its zero-shot advantage even when historical ERA5 data are partially or entirely unavailable.

We further examined the impact of historical weather availability while keeping single-cycle future GFS inputs, historical PV sequences, and test windows unchanged. Using the pretrained PVFM checkpoint, we masked the most recent 6, 24, or 120 hours of historical ERA5, as well as removing historical ERA5 entirely. We compared PVFM with pretrained TSFMs under the same conditions. The five-task average MAE results are presented below.

Table 3. Zero-Shot Performance under Historical Weather Availability Constraints.

Historical weatherPVFMBest TSFM
Complete ERA50.050230.07480
Recent 6h missing0.051980.08664
Recent 24h missing0.058250.07790
Recent 120h missing0.061750.07747
No historical ERA50.064710.07782

PVFM exhibits increasing errors as historical ERA5 becomes unavailable. Nevertheless, even without any historical ERA5, PVFM achieves 16.8% lower MAE than the best pretrained TSFM.

4. Adaptation to Operational Meteorological Data Availability through Post-Pretraining

To further improve PVFM's adaptability to practical forecasting conditions, we conducted 5,000 steps of post-pretraining starting from the pretrained PVFM checkpoint, using single-cycle GFS inputs from 457 seen sites. Historical ERA5 availability was varied during post-pretraining to accommodate different data availability conditions. Table 4. Impact of Post-Pretraining under Historical Weather Availability Constraints.

ModelComplete ERA5 (MAE / CRPS)No Historical ERA5 (MAE / CRPS)
PVFM0.05023 / 0.036290.06471 / 0.04687
PVFM (Post5k)0.04918 / 0.035210.05170 / 0.03708
  • Adaptation to missing historical weather. When historical ERA5 is entirely unavailable, post-pretraining reduces MAE and CRPS by 20.11% and 20.89%, respectively, compared with the original PVFM.
  • Adaptation to single-cycle GFS. With complete historical ERA5, mask-based post-pretraining reduces the five-task average MAE and CRPS by 2.09% and 2.96%, respectively, with MAE reductions of 2.90%–4.72% across the three hourly forecasting tasks.

Conclusion. Our original pretraining strategy was designed to learn transferable Weather-to-Power relationships through large-scale cross-site training with relatively complete meteorological inputs. The additional experiments demonstrate that PVFM retains its zero-shot advantages under realistic GFS availability constraints and remains effective even when historical ERA5 is unavailable. Furthermore, post-pretraining further improves its performance under these conditions, demonstrating its potential for adaptation to practical PV forecasting scenarios.

Revision. We have clarified the original GFS data organization in Appendix A.2 and added further evaluations and post-pretraining experiments under more realistic operational conditions in Appendix C.5.

Reviewer_qq9AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes PVFM, a 3.88M-parameter foundation model for zero-shot photovoltaic power forecasting.

The architecture is dual-branch:

  1. A historical encoder alternates temporal and cross-variable self-attention over patched PV capacity factor and six ERA5 meteorological variables.
  2. A decoder initializes latent future-PV tokens by projecting the encoder’s PV representation along the patch axis. These tokens then cross-attend to GFS future-weather tokens.
  3. Input tokens are augmented with solar-geometry embeddings (an elevation proxy plus hour-angle sin/cos), a site embedding (latitude, periodic longitude, UTC offset), and variable-identity embeddings.
  4. A non-crossing 11-quantile head is trained with pinball loss.

The model is pretrained on 746 sites (59.2M points, 1-hour and 15-minute resolution, 10 countries/regions). It is evaluated zero-shot on 99 held-out sites (79 for the 15-minute tasks) across five context/horizon settings, from 24→6 h up to 336→168 h. Baselines are per-station full-shot models (with FC/TC adaptations for future covariates) and six covariate-capable TSFMs. PVFM reports the best results on all five tasks, with MAE reductions of 10.2–24.3% and CRPS reductions of 10.0–29.5% (Table 1).

Additional experiments include:

  1. region holdout for Hong Kong and Australia (Sec. 6.2, Tables 7–8)
  2. architecture and embedding ablations (Fig. 5)
  3. a ramp-event evaluation (Table 10)
  4. a gated cross-attention satellite adapter fine-tuned on Himawari-9 imagery for 4-hour forecasting in Anhui (Table 11)
Critical strengths
  1. A large, heterogeneous PV corpus with a careful evaluation protocol. The corpus spans 845 sites across centralized, distributed, and rooftop installations on five continents (Table 2). The protocol uses chronological splits, explicit masks instead of imputation, station-equal aggregation, and a common Q9 grid for CRPS (App. B.1). This matters because PV forecasting research is fragmented across single-site, single-horizon setups (Fig. 1). A shared multi-region benchmark with a released code repository could be useful to the community regardless of how the modelling claims hold up.

  2. Consistent gains, including some genuinely out-of-region evidence. PVFM beats all six covariate-aware TSFMs on 14 of 15 headline task/metric cells in Table 1. More importantly, the region-holdout experiment (Sec. 6.2) shows it still beats TSFMs in Hong Kong when that region is excluded from pretraining, by up to 7.5% MAE (Table 7). This is the most convincing evidence that domain-specific pretraining with solar-geometry priors transfers in a way general-purpose TSFMs do not.

Critical weaknesses
  1. Possible future-weather leakage, and an in-region “zero-shot” setting that likely overstates transfer.

PVFM’s accuracy is almost flat across horizons:

MAE is 0.045 / 0.045 / 0.045 and R² is 0.830 / 0.829 / 0.824 for 6 h, 24 h, and 168 h ahead (Table 1). RMSE is 0.0970 / 0.0975 / 0.0977 (Table 6).

Every full-shot baseline degrades at 168 h. NWP cloud and irradiance skill falls sharply between day 1 and days 5–7, so a model whose 7-day-ahead error matches its 6-hour-ahead error is surprising. The paper says only that GFS covariates are “aligned with the PV sequence according to their valid timestamps” (App. A.2). It never states:

which GFS cycle is used for a given forecast origin what lead time each valid time corresponds to whether publication latency is respected how GFS’s 3-hourly output beyond 120 h is handled

If covariates are stitched from short-lead forecasts of successive cycles, the model effectively sees near-analysis future weather. That would invalidate the long-horizon comparisons and inflate all of them.

Separately, the 99 “unseen” sites come from the same regions, and largely the same time spans, as the 746 pretraining sites. For example, Anhui contributes 218 pretraining sites versus 17 evaluation sites, all from 2023–2024 (Table 2). It is not stated whether pretraining sites were cut at the same calendar boundary as target sites’ test periods. If they were not, the model has seen PV output from nearby plants under the same weather events it is tested on.

Consistent with this concern, the gains shrink when a region is truly held out (Sec. 6.2, Table 7):

In Australia, PVFM has the best MAE on only 2 of 5 tasks. It trails iTransformer(TC) by 9.7% and 4.3% on the two 15-minute tasks. It is behind TimesFM-3 at 336→168.

This matters because the headline claims (10–24% MAE reduction, “consistently outperforms full-shot models”) rest entirely on the in-region setting.

  1. Missing standard solar-forecasting baselines and evaluation conventions, plus unequal information across methods.

There is no smart/clear-sky persistence, no physical NWP→power model, and no gradient-boosted model on NWP irradiance plus solar zenith. These are standard and often strong references in this literature, including the cited Zhang et al. (2026). No forecast skill scores are reported.

Metrics appear to be computed over all timesteps, including night. PVFM receives an explicit solar-elevation feature, so night zeros come for free. The full-shot baselines and TSFMs get neither solar geometry nor site coordinates (Table 3). Including night inflates R² and narrows the meaningful comparison, so part of the reported margin may come from this information asymmetry rather than from the architecture.

The fairer comparison is baselines trained on the same pooled corpus with the same solar features. App. C.3.2 attempts this but reports only a normalized radar chart (Fig. 8) with no table or numbers.

Finally, PVFM and the zero-shot results are reported from a single run with no variance. Seed variability is given only for the full-shot baselines (App. C.6). Several ablation effects are within 0–1% (Fig. 5), which cannot be interpreted without variance.

Smaller issues (not decisive):

  1. Satellite results. Table 11 and Fig. 6(c–d) disagree: fine-tuned RMSE is 0.0842 vs. 0.0839, and R² is 0.8628 vs. 0.8642. The abstract’s 9.1% matches the figure, not the table. Fine-tuned PVFM’s MAE of 0.0390 is worse than FusionSF (0.0377) and CrossViViT (0.0383), so “surpasses FusionSF” holds only for RMSE and R².
  2. Backbone mismatch. The satellite experiments appear to use a 16.2M-parameter backbone with d=256 (Table 5), not the 3.88M headline model (Table 9).
  3. Undisclosed baseline. CrossViViT appears in Tables 5 and 11 but not in the baseline list (Sec. 5.2).
  4. Naming. “PVFM-Tiny” appears in Fig. 8 without definition.
  5. Typos. “Main reulsts” (Sec. 6.2). Sec. 6.1 says evaluation uses MAE/RMSE/R², but Table 1 reports R²/MAE/CRPS.
  6. Scaling. Fig. 7(b) shows zero-shot R² decreasing with model size, which deserves discussion given the foundation-model framing.
Questions for authors

Questions that could change my assessment:

  1. GFS construction. For a forecast issued at time t with horizon H, which GFS cycle supplies the covariates, at what lead times, and is the cycle’s publication delay respected? How are 3-hourly GFS steps beyond 120 h interpolated? Can you report PVFM’s error as a function of lead time, for example days 1–7 within the 336→168 task? Showing the expected degradation would largely resolve my main concern.

  2. Temporal overlap. Were pretraining sites truncated so that no pretraining data overlaps the calendar period of the target sites’ test segments? If not, can you report the main table under a strict temporal cutoff?

  3. Physical/statistical baselines. Can you add smart persistence and a GBDT model on GFS irradiance plus solar zenith? Can you also report daytime-only metrics (e.g., solar elevation > 0) or skill scores relative to persistence?

  4. Pooled baselines. Can you provide a table, not just the radar chart, for baselines pretrained on the same corpus (App. C.3.2)? Did those baselines receive the same solar and site embeddings? If not, how much of PVFM’s gain survives when they do?

  5. Variance. Can you report variance across pretraining seeds for PVFM on Table 1 and on the ablations in Fig. 5?

  6. Timekeeping. Timestamps are in site-local time with a fixed UTC offset z_s (Eq. 33). How is daylight saving time handled for the UK, US, and Australian sites? An unhandled one-hour shift would misalign the solar features for half the year.

  7. Data availability. Several pretraining sources (Anhui, Guangxi, and other Chinese “project industrial PV records”) appear proprietary. Which parts of the corpus will be released? The reproducibility statement implies full reproducibility.

8 comments
AuthorsOpenReview ↗

Response to Reviewer qq9A 1/7

W1 & Q1: How are GFS forecast cycles selected, and how does error vary with lead time?

Weakness 1. “If covariates are stitched from short-lead forecasts of successive cycles, ... the model effectively sees near-analysis future weather. That would invalidate the long-horizon comparisons ...”

Question 1. “... which GFS cycle supplies the covariates, at what lead times, and is the cycle’s publication delay respected? How are 3-hourly GFS steps beyond 120 h interpolated? ... report PVFM’s error as a function of lead time, for example days 1–7 within the 336→168 task ...”

Response. We thank the reviewer for raising concerns about GFS forecast initialization times, meteorological data availability, and long-horizon forecasting performance. Our original setup was designed to support large-scale cross-site pretraining using relatively complete and accurate meteorological inputs, enabling PVFM to learn transferable Weather-to-Power relationships across diverse conditions. However, we acknowledge that the original valid-time-aligned GFS construction did not enforce forecast-origin availability, representing an idealized meteorological-input setting rather than a fully operational forecasting protocol. This limitation is particularly relevant for longer forecasting horizons. To assess the robustness of our results, we conducted an additional forecast-origin-constrained evaluation using single-cycle GFS forecasts with an explicit publication delay.

Clarification: GFS Data Construction and Forecast Availability

The original experiments used continuous meteorological sequences aligned by valid time. Future meteorological covariates were extracted according to the valid timestamps of the PV forecasting targets, rather than selecting a GFS cycle available at each forecast origin. The same GFS data construction and temporal alignment protocol was applied to PVFM and all TSFM and full-shot baselines using GFS inputs. These sequences were constructed from two data sources:

  • On or after March 23, 2021: We obtained continuous hourly historical GFS forecast sequences through Open-Meteo.
  • On or before March 22, 2021: We retrieved eight forecast lead times (+3, +6, .., +24 h) from the daily 00 UTC GFS cycles archived by GDEX, organized them by valid time, and linearly interpolated them to hourly resolution.

Further Analysis: Forecasting Performance Under Realistic GFS Constraints

We conducted additional experiments incorporating GFS initialization times and publication delays to examine how changes in meteorological inputs affect forecasting accuracy and whether PVFM retains its zero-shot advantages under more realistic forecasting conditions. Due to the limited rebuttal period and GFS download throughput, we sampled complete GFS forecast cycles across the original test period.

1. Additional GFS construction with forecast-origin constraints.

  • Additional GFS data collection. We retrieved complete GFS forecast cycles from NOAA AWS and GDEX, using 3-hourly forecasts up to +240 h and interpolating them to hourly inputs within each cycle.
  • Forecast-cycle selection and availability constraints. Assuming a uniform 6-hour publication delay, we selected the latest eligible cycle among the downloaded GFS forecasts at each PV forecast origin, requiring it to cover the full prediction horizon. All future meteorological inputs were extracted from this single cycle, without cross-cycle stitching.
AuthorsOpenReview ↗

Response to Reviewer qq9A 2/7

W1 & Q1 (continued)

2. Impact of Replacing GFS Inputs on PVFM Forecasting Accuracy

(1) Performance changes and zero-shot comparison.

Finding-1. Hourly forecasting tasks are more affected by single-cycle GFS inputs, particularly at longer horizons. Keeping the pretrained checkpoint, test windows, and historical inputs unchanged, we replaced only the future GFS inputs with single-cycle forecasts. MAE increased by 2.6%, 9.1%, and 24.2% for the three hourly tasks (6 h, 24 h, and 168 h), respectively, compared with only 1.1% and 1.5% for the two 15-minute tasks (4 h and 6 h).

Finding-2. PVFM retains its daytime zero-shot advantage under identical single-cycle GFS inputs. We compared PVFM with five pretrained TSFMs and three target-site full-shot models using their original checkpoints and identical single-cycle GFS inputs. PVFM achieves the best five-task average performance across all five metrics, reducing MAE by 28.8% and 13.4% relative to the best TSFM and full-shot baseline, respectively.

Table 1. Daytime Zero-Shot Performance under Single-Cycle GFS Inputs. All models were evaluated on identical daytime targets (solar elevation > 0°). Best TSFM denotes the best pretrained TSFM for each metric, while Best Full-shot corresponds to iTransformer-TC.

ModelR² ↑MAE ↓RMSE ↓CRPS ↓AQL ↓
PVFM0.63000.09780.14460.07050.0382
Best TSFM0.31380.13730.19590.10190.0552
Best Full-shot0.52870.11300.15980.08890.0477
vs. Best TSFM+0.316228.8%26.2%30.7%30.7%
vs. Best Full-shot+0.101313.4%9.5%20.6%19.8%

(2) Lead-time-dependent errors within the 336→168 task.

Finding-3. Forecasting errors generally increase with lead time over the seven-day horizon, while PVFM consistently maintains lower errors than the strongest pretrained TSFM. We further compared PVFM with the best-performing pretrained TSFM (Chronos-2) across forecast Days 1–7 using identical single-cycle GFS inputs and test windows.

Table 2. Daily Forecasting MAE over the 168-Hour Horizon.

Forecast dayPVFMChronos-2MAE reduction
Day 10.04910.058816.6%
Day 20.04990.059115.6%
Day 30.05690.066013.7%
Day 40.05630.068617.9%
Day 50.05970.068813.2%
Day 60.06100.071414.6%
Day 70.06120.070112.7%

Both models exhibit generally increasing errors with longer forecast lead times, consistent with the expected decline in meteorological forecast accuracy. Nevertheless, PVFM maintains lower MAE than Chronos-2 on every forecast day, with reductions of 12.7%–17.9%.

3. PVFM Post-Pretraining

To further adapt PVFM to realistic GFS availability constraints, we conducted 5,000 steps of post-pretraining using single-cycle GFS inputs from 457 seen sites, starting from the original PVFM checkpoint. We focused on the three hourly forecasting tasks, which were more affected by the transition to single-cycle GFS inputs.

Table 3. Effect of Post-Pretraining on Hourly Forecasting Performance.

TaskOriginal PVFMPVFM-Post 5kMAE change
1h:24→60.049530.04719−4.72%
1h:72→240.052990.05054−4.62%
1h:336→1680.056300.05467−2.90%

With 5,000 steps of post-pretraining, PVFM achieves MAE reductions of 4.72%, 4.62%, and 2.90% across the three hourly forecasting tasks, respectively. These results suggest that limited post-pretraining can further improve forecasting performance under realistic GFS availability constraints.

Conclusion. Our original setting was designed to support large-scale cross-site pretraining with relatively accurate and complete meteorological inputs, enabling PVFM to learn transferable Weather-to-Power relationships. The additional experiments demonstrate that PVFM maintains its zero-shot advantage over the strongest baselines under realistic GFS availability constraints. Furthermore, we conducted preliminary post-pretraining on the more affected hourly forecasting tasks, achieving further performance improvements and demonstrating the potential to adapt PVFM to operational forecasting conditions.

Revision. We have clarified the original GFS data organization in Appendix A.2 and added further evaluations and post-pretraining experiments under more realistic operational conditions in Appendix C.5.

AuthorsOpenReview ↗

Response to Reviewer qq9A 3/7

W2 & Q2: Does temporal overlap affect the zero-shot evaluation, and how well does PVFM transfer to held-out regions?

Weakness 2. “... the 99 ‘unseen’ sites come from the same regions, and largely the same time spans, as the 746 pretraining sites. ... In Australia, PVFM has the best MAE on only 2 of 5 tasks. ... the headline claims ... rest entirely on the in-region setting.”

Question 2. “Were pretraining sites truncated so that no pretraining data overlaps the calendar period of the target sites’ test segments? If not, can you report the main table under a strict temporal cutoff?”

Response. We thank the reviewer for raising concerns about temporal overlap and cross-region generalization. We conducted an additional evaluation excluding test windows overlapping with same-region pretraining data to assess its impact on zero-shot performance.

  • Filtering rule. We excluded test windows whose forecast-target timestamps overlapped with any same-region pretraining data, considering both hourly and 15-minute records.
  • Baseline evaluation. All models were evaluated on identical retained test windows using saved predictions or inference with their original checkpoints. No models were retrained.

Table 4. Performance after Excluding Test Windows with Same-Region Pretraining Overlap. Each entry reports R² (↑) / MAE (↓) / CRPS (↓), with station-equal aggregation.

SettingModel1 h: C24→H61 h: C72→H241 h: C336→H16815 min: C64→H1615 min: C96→H24
Full-shotiTransformer-TC0.778 / 0.053 / 0.0510.772 / 0.057 / 0.0530.637 / 0.082 / 0.0710.768 / 0.046 / 0.0430.796 / 0.047 / 0.043
Zero-shotChronos-20.596 / 0.077 / 0.0580.744 / 0.056 / 0.0410.736 / 0.057 / 0.0400.293 / 0.077 / 0.0590.566 / 0.068 / 0.050
Zero-shotPVFM0.816 / 0.042 / 0.0300.818 / 0.042 / 0.0300.807 / 0.043 / 0.0310.758 / 0.039 / 0.0290.806 / 0.040 / 0.029

After excluding test windows overlapping with same-region pretraining data, PVFM maintains the lowest MAE and CRPS across all five tasks among the compared models.

Revision. We have expanded Appendix C.1 to include the complete evaluation results for all full-shot and zero-shot baselines under the same-region temporal overlap-excluded setting, reporting R², MAE, and CRPS across all five forecasting tasks (Table 8).


W3, W4 & Q3: How does PVFM compare with standard solar-forecasting baselines under daytime evaluation?

Weakness 3. “There is no smart/clear-sky persistence, no physical NWP→power model, and no gradient-boosted model ... No forecast skill scores are reported.”

Weakness 4. “Metrics appear to be computed ... including night. PVFM receives an explicit solar-elevation feature ... The full-shot baselines and TSFMs get neither solar geometry nor site coordinates ... part of the reported margin may come from this information asymmetry rather than from the architecture.”

Question 3. “... add smart persistence and a GBDT model on GFS irradiance plus solar zenith? ... report daytime-only metrics (e.g., solar elevation > 0) or skill scores relative to persistence?”

Response. We thank the reviewer for these suggestions. We have supplemented the main results with four conventional solar-forecasting baselines: persistence, smart persistence, a physical NWP-to-power model (PVWatts), and GBDT using GFS irradiance and solar zenith. We also reevaluated all models on identical daytime targets (solar elevation > 0°). The daytime MAE results are presented below.

Table 5. Daytime MAE Comparison with Conventional Solar-Forecasting Baselines.

Model1h/6h1h/24h1h/168h15min/4h15min/6h
Persistence0.2600.3040.3030.1740.218
Smart persistence0.1590.1980.2120.1110.146
Physical (PVWatts)0.1750.1800.1710.1730.174
GBDT (GFS + zenith)0.1280.1300.1260.1190.131
Best existing full-shot0.1010.1050.1310.0860.092
Best TSFM0.1130.1100.1070.1470.133
PVFM0.0870.0870.0880.0820.086

PVFM achieves the lowest daytime MAE across all five tasks, reducing MAE by 4.5–18.0% relative to the strongest competing method. It also achieves a station-averaged RMSE skill score of 46.9–64.5% relative to ordinary persistence. These results demonstrate that PVFM maintains its forecasting advantages when nighttime observations are excluded and conventional solar-forecasting baselines are included.

Revision. We added four conventional solar-forecasting baselines and a daytime evaluation (solar elevation > 0°) in Appendix C.1, with results reported in Table 7.

Reviewer_RndZOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Authors presents a model for photovoltaic power forecasting. The model is looking at historical timeseries data and weather and also at satellite images to predict the amount of power generated. Overall the results show good performance in MAE reduction when compared to several other baselines.

I would say that labelling it as foundation model stretches the current industry definition; The satellite images for instance are just used as a fine-tuning adapter. I haven't seen an Spatial-Temporal baseline (something similar to this used for solar irradiance, ex. SunCast: Solar Irradiance Nowcasting from Geosynchronous Satellite Data). The spatial-temporal baseline can help at tracking localized cloud movement (the primary driver of variance in 15 minutes forecasting windows).

comments:

  1. I would recommend to soften a little bit the "foundational model" branding to something more precise. If retained please provide additional details;
  2. For ultra-short term prediction NWP models would be too slow to update, try to consider ingesting spatial temporal models, like the Suncast.
  3. Give additional info on why for certain target-region we have lower performance (ref. specific data distribution changes).
Critical strengths

Great results, it's a very practical paper, and several baselines were used in the evaluation section.

Critical weaknesses

I think the main weakness is relying too much on the meteorological models (which might be not that fresh always), which make the usage of this solution difficult in real-time. Now the proposed method is to compensate by introducing satellite adapter, I would strong suggest to further investigate this approach, and look at the reference I recommended.

Questions for authors

How sensitive is PVFM to the misaligned weather data without satellite feedback?

4 comments
AuthorsOpenReview ↗

Response to Reviewer RndZ 1/3

C1. Foundation-model positioning and justification

Comment 1. "I would recommend to soften a little bit the "foundational model" branding to something more precise. If retained please provide additional details;"

Response. We sincerely thank the reviewer for this suggestion. We would like to clarify that PVFM is specifically designed as a domain-specific foundation model for photovoltaic power forecasting. This positioning is supported by its large-scale cross-site pretraining, zero-shot forecasting capabilities on unseen sites, and adaptability to downstream forecasting scenarios, as detailed below.

1. Definition and scope. Foundation models for energy forecasting are generally characterized by large-scale pretraining, transferable representations, and adaptation across diverse forecasting scenarios, as discussed by Ferdaus et al. (2026)[1]. This paradigm is reflected in time-series forecasting models such as Chronos[2] and TimesFM[3], as well as domain-specific wind power forecasting models including DeepWind[4], WindFM[5], and Tyan-WP[6]. PVFM follows this paradigm by learning transferable weather–power relationships from diverse photovoltaic data to support forecasting across previously unseen sites and different forecasting scenarios.

2. Evidence of foundation-model capabilities. PVFM follows the above foundation-model paradigm in photovoltaic power forecasting, as demonstrated by:

  • Large-scale pretraining: Pretrained on PV–weather time series from 746 sites across multiple countries and regions, covering diverse PV system types and environmental conditions.
  • Forecasting generalization: A single pretrained model supports two temporal resolutions (hourly and 15-minute) and flexible forecasting horizons of up to 168 time steps.
  • Zero-shot transfer: Evaluated on 99 unseen sites and geographically held-out regions without target-site parameter updates, with comparisons against site-specific forecasting models.
  • Downstream adaptation: PVFM supports target-site fine-tuning and multimodal adaptation with satellite imagery, demonstrating its adaptability to different forecasting settings and input modalities.

These capabilities support PVFM's positioning as a domain-specific foundation model for photovoltaic power forecasting. Our focus is on forecasting as a fundamental task for renewable energy operation and dispatch, through which PVFM learns transferable weather–power relationships across diverse PV sites and environmental conditions.

[1] Ferdaus, M. M., Dam, T., Sarkar, M. R., Uddin, M. & Anavatti, S. G. Foundation models for clean energy forecasting: A comprehensive review. Renewable and Sustainable Energy Reviews, 226, 116452 (2026).

[2] Ansari, A. F. et al. Chronos: Learning the language of time series. Transactions on Machine Learning Research (2024).

[3] Das, A., Kong, W., Sen, R. & Zhou, Y. A decoder-only foundation model for time-series forecasting. In Proceedings of the 41st International Conference on Machine Learning, PMLR, 235, 10148–10167 (2024).

[4] Wang, H., Zhou, T., Jia, C., Liu, Y. & Wang, L. DeepWind: A foundation model for zero-shot wind power forecasting. Energy, 360, 141793 (2026).

[5] Fan, H. et al. WindFM: An open-source foundation model for zero-shot wind power forecasting. arXiv preprint arXiv:2509.06311 (2025).

[6] Huang, J. et al. Tyan-WP: A wind power foundation model for ultra-short-term probabilistic forecasting. arXiv preprint arXiv:2606.08630 (2026).

AuthorsOpenReview ↗

Response to Reviewer RndZ 2/3

C2 & W1. NWP Timeliness and Integration of Spatiotemporal Models for Ultra-Short-Term Forecasting

Comment 2. “... NWP models would be too slow to update, ... consider ingesting spatial temporal models, like the Suncast.”

Weakness 1. “... relying too much on the meteorological models ... difficult in real-time. ... introducing satellite adapter, I would strong suggest to further investigate this approach ...”

Response. We sincerely thank the reviewer for this valuable suggestion. The update frequency of meteorological forecasts is indeed an important consideration in operational PV forecasting, particularly for ultra-short-term predictions. Spatiotemporal forecasting models such as SunCast offer a promising alternative for providing more frequently updated meteorological inputs to PVFM.

Additional experiment. We implemented SunCast using Himawari satellite-derived irradiance data over Anhui, China. Its predicted irradiance replaced only the future NWP shortwave irradiance input, while the remaining inputs and pretrained PVFM checkpoint were unchanged. We evaluated zero-shot performance on 17 unseen PV sites under the 16-hour-context/4-hour-horizon forecasting task.

Table 1. Zero-Shot Performance with SunCast-Derived Irradiance Inputs.

Meteorological inputRMSE ↓MAE ↓R² ↑
Original NWP0.1157920.0594620.769591
SunCast-derived irradiance0.1163630.0612060.767311
Relative change+0.49%+2.93%−0.30%

Replacing the NWP irradiance input with SunCast predictions increased RMSE and MAE by only 0.49% and 2.93%, respectively, while R² decreased by 0.30%. These preliminary results suggest that satellite-based spatiotemporal models could offer a promising alternative to NWP inputs for more timely ultra-short-term PV forecasting.

Revision. We have expanded the Related Work section to include satellite-based spatiotemporal models such as SunCast, highlighting their potential to complement conventional NWP products with satellite-derived irradiance forecasts.

AuthorsOpenReview ↗

Response to Reviewer RndZ 3/3

C3: Why does PVFM perform worse in certain target regions?

Comment 3. "Give additional info on why for certain target-region we have lower performance (ref. specific data distribution changes)."

Response. We thank the reviewer for this suggestion. In our region-holdout experiments, we observed relatively weaker zero-shot forecasting performance in Australia compared with Hong Kong. We therefore selected Australia for further investigation, focusing on regional data distribution characteristics and their relationship with forecasting errors.

1. Regional distribution differences. We compared the 569 non-Australian sites in the 15-minute pretraining subset with 11 Australian target sites using station-equal statistics. We further matched irradiance, solar elevation, and temperature to examine differences in power generation under similar meteorological conditions.

Table 2. Regional Data Distribution Comparison between Australian and Non-Australian PV Sites.

CharacteristicNon-AU pretrainingAU test
Mean irradiance (W/m², GHI > 50)413.24552.11
High-output frequency (CF > 0.5)27.91%45.95%
Large 15-min power changes (ΔCF > 0.15)7.84%15.60%
High-output plateau frequency1.46%14.63%
Mean CF under matched conditions0.42680.4697

Australian sites exhibit nearly twice the frequency of large power changes and approximately 10 times the frequency of high-output plateaus. Even under matched meteorological conditions, their mean CF remains 10.1% higher than that of non-Australian sites, indicating substantial differences in regional generation characteristics.

2. Relationship with forecasting errors. To further investigate the relationship between regional data distribution and forecasting errors, we evaluated two PVFM variants pretrained with and without Australian data, denoted as PVFM (+AU) and PVFM (−AU), on the same 11 unseen Australian sites excluded from both pretraining datasets.

Table 3. Relative MAE Changes of PVFM (−AU) Compared with PVFM (+AU) under Low- and High-Output Conditions (GHI > 50 W/m²). Results are aggregated equally across stations.

Output regime4h ΔMAE6h ΔMAE
Low output (CF ≤ 0.5)−2.8%+3.2%
High output (CF > 0.5)+62.4%+40.1%

Compared with PVFM (+AU), PVFM (−AU) exhibits 62.4% and 40.1% higher MAE under high-output conditions for the 4h and 6h tasks, respectively, while low-output errors change by only −2.8% and +3.2%.

Conclusion. These results suggest that distinctive regional generation characteristics, particularly under high-output conditions, may limit cross-region zero-shot generalization. Such differences may be associated with PV system configurations and operational constraints, motivating future integration of such information to better distinguish region-specific characteristics from transferable patterns across PV sites.


Q1: How sensitive is PVFM to weather-data misalignment without satellite feedback?

Question 1. "How sensitive is PVFM to the misaligned weather data without satellite feedback?"

Response. Thank you for the question. We evaluated PVFM's sensitivity to future NWP timing errors by shifting all six NWP variables by ±1h and ±2h across five forecasting tasks. The original pretrained model was evaluated without satellite feedback or fine-tuning, while all other inputs remained unchanged.

Table 4. Sensitivity of Zero-Shot Forecasting to NWP Temporal Misalignment. Values represent relative changes in station-equal MAE compared with the unshifted NWP inputs.

NWP offset1h / 6h1h / 24h1h / 168h15min / 4h15min / 6h
−2h+21.13%+18.80%+17.79%+16.32%+20.44%
−1h+7.12%+6.09%+5.45%+7.10%+7.99%
0h0%0%0%0%0%
+1h+5.10%+5.31%+5.46%+0.68%+3.24%
+2h+15.76%+17.00%+17.50%+5.34%+11.17%

Our findings are twofold:

  • Sensitivity to shift magnitude. Larger temporal shifts generally lead to greater forecasting errors. Even a 1-hour shift increases MAE by 0.68%–7.99%, while 2-hour shifts result in increases of 5.34%–21.13%.

  • Asymmetric effects of shift direction. Negative shifts consistently cause larger MAE increases than positive shifts across all five tasks. This asymmetry is particularly evident in 15-minute forecasting, where a 2-hour negative shift increases MAE by 16.32% and 20.44%, compared with 5.34% and 11.17% for positive shifts.

These findings highlight the importance of accurate temporal alignment of future meteorological inputs, including both shift magnitude and direction.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer RndZ,

Thank you for your time and effort in reviewing our paper. Your comments are valuable in helping us improve the quality of our paper. As the discussion deadline approaches, we kindly invite you to review our rebuttal and let us know whether your concerns have been adequately addressed.

We greatly appreciate your time and consideration.

Regards from the authors.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We sincerely thank all reviewers for their constructive feedback. During the discussion period, we conducted additional experiments focusing on two major issues: realistic meteorological data availability and the fairness and robustness of experimental comparisons. Below, we summarize our findings and revisions.

1. PVFM retains its forecasting advantages under realistic meteorological data availability constraints

Reviewer comments: RndZ (C2/W1); qq9A (W1/Q1); 3Vem (W3/Q2).

Manuscript revisions: Appendix A.2 and C.5 (Tables 15–16).

Our original study focused on learning transferable weather-to-power relationships across diverse PV sites under an idealized meteorological-input setting. All applicable models used identical meteorological data and temporal alignment to ensure fair comparisons. We also considered ERA5 availability limitations in our original satellite adaptation experiments. Following the reviewers' suggestions, we further investigated PVFM's practical applicability by jointly considering historical and future weather availability, an issue that remains relatively underexplored in existing PV forecasting studies.

EvaluationHistorical weatherFuture weather (NWP)Key findings
Original study(1) Complete historical ERA5 inputs; (2) Satellite adaptation without ERA5 for ultra-short-term forecasting.Identical GFS inputs aligned by valid time across models.PVFM outperforms competing baselines, reducing MAE by 10.2–24.3% and CRPS by 10.0–29.5% across five tasks.
Reviewer-requested experimentsSimulated ERA5 availability delays (6, 24, and 120 h) and complete unavailability.(1) Single-cycle GFS with an assumed 6 h publication delay; (2) SunCast-derived irradiance as an alternative input.20.46% lower MAE than the best baseline with single-cycle GFS; 16.8% lower MAE than the best TSFM when all models operate without ERA5. Compatibility with alternative weather sources (SunCast: +0.49% RMSE).
Further explorationPost-pretraining with historical weather masking.Post-pretraining with single-cycle GFS inputs.With only 5,000 post-pretraining steps, PVFM further reduces MAE by 2.90–4.72% across three hourly tasks with single-cycle GFS and by 20.11% under ERA5-free conditions.

Despite time and download constraints, we retrieved as many complete single-cycle GFS forecasts as possible, covering all 99 unseen sites. Importantly, PVFM outperforms all evaluated baselines on average under both the original idealized setting and the more realistic meteorological data availability constraints.

2. PVFM's performance gains persist under more strictly controlled comparisons

Reviewer comments: qq9A (W2–W6/Q2–Q5); 3Vem (W1/Q1).

Manuscript revisions: Appendix C.1 (Tables 7–8) and C.3.2 (Table 12).

  • Controlled architectural comparisons. Beyond pretraining competing architectures on the same corpus, we further incorporated PVFM's solar-aware and site information into these baselines through their respective embedding mechanisms. PVFM retains a 3.10% average MAE advantage (11.25% at 168 h). These results highlight the effectiveness of our embeddings and suggest the benefits of separating historical and future meteorological inputs for long-horizon forecasting.
  • Temporal generalization and additional baselines. After excluding test windows overlapping with same-region pretraining data, PVFM retains the lowest MAE and CRPS across all five tasks. Additional daytime evaluations incorporating four conventional solar-forecasting baselines show MAE reductions of 4.5–18.0% relative to the strongest competing methods.
  • Controlled satellite adaptation. Additional controlled experiments using identical updated GFS inputs demonstrate the incremental benefits of incorporating satellite imagery.
  • Experimental stability. We repeated PVFM pretraining and all four architectural ablations across three independent seeds. The results support the effectiveness and stability of the proposed architecture.

3. Other manuscript revisions

In addition to the experiments above, we clarified PVFM’s positioning, model configurations, and experimental protocols, expanded the related work and discussion, and corrected reporting inaccuracies in the appendix. All revisions are highlighted in blue in the revised manuscript.

We believe we have addressed the reviewers’ concerns through our individual responses and additional experiments. We sincerely appreciate their constructive feedback, which has strengthened the experimental rigor and practical relevance of our work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.