acceptodds
Under review as a conference paper at ICLR 2027

W2P-FM: Learning Transferable Weather-to-Power Representations for Wind Power Forecasting Across Sites and Tasks

Abstract

As wind power expands globally, accurate forecasting is critical for managing weather-driven generation variability and enabling its reliable integration into modern energy systems. Yet, continued wind-farm deployment and diverse forecasting demand across horizons and temporal granularities expose a fundamental limitation of conventional data-driven models, which are typically optimized on specific sites and tasks. Meeting these requirements calls for transferable modeling of weather-driven power dynamics while addressing three key challenges: aligning multi-source weather and power data, capturing cross-site heterogeneity, and supporting diverse forecasting tasks within a unified framework. To tackle these challenges, we propose W2P-FM, a foundation model for wind power forecasting pretrained on more than 20 million time points. W2P-FM constructs a uniform temporal-variable representation by using MeteoRouter to adaptively fuse numerical weather predictions and granularity-aware patching to align power and meteorological sequences. It further incorporates rich static site context and recent response observations to characterize site-specific patterns. Finally, weather-to-power representations are learned through intra-patch cross-variable interactions and local multi-scale inter-patch aggregation, with task-specific heads supporting different forecasting settings. Experiments on 33 real-world datasets comprising over 1,000 site-level wind power series across five continents show that W2P-FM consistently outperforms state-of-the-art forecasting models, reducing MSE by 12.3% in full-shot and 48.9% in zero-shot settings. Frozen-backbone adaptation reduces MSE by 17.2% in cross-granularity forecasting and CRPS by 15.8% in probabilistic forecasting relative to the strongest task-specific baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
wind power forecasting, foundation models, time series, zero-shot generalization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3EP8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a foundation model for wind power forecasting. It combines fusion of multiple weather forecasts, time aligned patching, static site features, and historical weather and power observations. Experiments cover 1,012 wind power series from 33 datasets, including tests on unseen sites, unseen regions, probabilistic forecasting, and finer time resolutions.

Critical strengths
  1. The paper addresses a consequential problem on a large heterogeneous corpus , and its chronological splits and explicit cross-region holdout are more informative than random-window benchmarks.
  2. The paper maps its design to identifiable forecasting challenges, and ablations support the key choices.
Critical weaknesses
  1. The paper does not separate its gains from pretraining scale and input access. W2P-FM is pretrained on 938 sites and fuses three NWP sources, while the baselines train on each dataset with a single source. The 12.3% improvement is therefore not attributable to the architecture alone.

  2. The zero shot setting still uses each site's recent power and weather history at inference, as shown in Equation 1 and Section 4.2. It tests transfer without parameter fitting, not the cold start case motivating the paper. The preprocessing details in Appendix C.1 are unclear.

  3. The paper reports no per seed variation or confidence intervals despite averaging over three seeds. This matters for small margins such as the 1.77% improvement for North America at 24 hours, as reported in Table 1. Geographic coverage is also uneven, with no North American sites in the holdout and no African sites in the corpus.

Questions for authors
  1. Could you add a controlled comparison with matched pretraining sites, horizon mixture, and NWP sources against a strong pooled data baseline? If the baseline cannot natively accept multiple NWP products, could you evaluate a simple matched input fusion variant? This would help separate corpus scale, additional weather information, and architecture contributions.

  2. What target site history is used at inference for held out sites? How do results change with shorter history or without Meteo Power Encoding? How are anomalous power readings identified, and can short gap interpolation be guaranteed not to use future observations or cross split boundaries?

  3. Could you report per seed results or confidence intervals for the main full shot and zero shot comparisons, ideally with a paired site level analysis? Could you also clarify how the missing North American and African sites bound the global generalization claim?

7 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1 of 4)

Dear Reviewer 3EP8,

We sincerely thank you for your thoughtful review. We are grateful that you acknowledge the informative evaluation on a large heterogeneous corpus and the challenge-driven design supported by ablations.

W1 & Q1: Controlled Comparison under Matched Data and NWP Inputs

Following your suggestion, we conduct controlled comparisons against strong pooled baselines with matched pretraining sites, horizon mixture, and NWP inputs, and find that W2P-FM-NS still achieves the lowest MAE and MSE across all three horizons, reducing average errors by 5.76%–6.56% relative to the strongest baseline. With corpus scale and weather information matched, these results demonstrate a clear advantage of its weather-to-power representation learning architecture. Moreover, comparing W2P-FM-NS under different NWP inputs shows that MeteoRouter effectively exploits the complementary information in multi-source NWP. The related experiments and analyses have been added to Appendix C.13.

(1) Controlled comparison experiment

  • All models are trained from scratch on the same 938 pretraining sites, chronological partitions, and training samples, without static site context.
  • Crossformer, PatchTST-GRU, and W2P-FM-NS are jointly trained across the context–horizon configurations (24, 6), (72, 24), and (72, 168); LightGBM, which does not natively support variable-length configurations, is trained separately for each horizon.
  • All models receive either GFS or Uniform NWP input (the equally weighted average of the available sources), the latter serving as the simple matched input-fusion variant you suggested. Results (MAE/MSE) are averaged over three random seeds.
NWP inputHorizonW2P-FM-NSCrossformerPatchTST-GRULightGBM
Uniform6h0.0999/0.02180.1070/0.02350.1098/0.02430.1143/0.0251
Uniform24h0.1156/0.02870.1255/0.03080.1307/0.03250.1390/0.0348
Uniform168h0.1564/0.04750.1649/0.05040.1657/0.05010.1753/0.0530
GFS6h0.1040/0.02300.1104/0.02430.1107/0.02470.1173/0.0269
GFS24h0.1229/0.03030.1306/0.03330.1343/0.03420.1420/0.0376
GFS168h0.1568/0.04790.1684/0.05050.1661/0.05040.1762/0.0537

(2) Contribution of additional weather information

These experiments also isolate the contribution of additional weather information. For W2P-FM-NS, the full-shot MSE decreases consistently from GFS to Uniform input and further to MeteoRouter-based fusion (W2P-FM-NS in Table 1). This indicates that multi-source NWP provides complementary information, which MeteoRouter exploits more effectively than uniform averaging, consistent with the corresponding ablation of W2P-FM (Table 8).

HorizonGFSUniformMeteoRouter
6h0.02300.02180.0210
24h0.03030.02870.0279
168h0.04790.04750.0469

W2: Zero-Shot Setting

In our zero-shot setting, each target site's historical information is limited to the short context available immediately before each forecast origin (24 or 72 hours), without any fine-tuning on the target sites. This setting corresponds to the scenario motivated in the Introduction (Lines 44–47), where newly commissioned sites lack sufficient data for model training but can provide recent observations. It is also consistent with existing zero-shot forecasting studies, including WindFM (Fan et al., 2025) and Chronos-2 (Ansari et al., 2025). We have clarified this evaluation setting in Appendix C.1.

Q2: Shorter Historical Context and MPE Ablation

To address your questions, we conduct two additional experiments with shortened historical context and without MPE, and find that even with only one-quarter of the original context, W2P-FM still substantially outperforms all zero-shot baselines that use the full context, while removing MPE consistently degrades performance, particularly for unseen sites and longer horizons. We have added both analyses to Appendix C.14.

(1) Sensitivity to historical context length. Using the original pretrained checkpoints on the 74 cross-site zero-shot test sites without retraining, we shorten both historical power and meteorological observations to one-half and one-quarter of the original length. Shortened inputs are left-padded with zeros and marked as invalid positions, preserving the original input shape.

HorizonContextMSEMSE IncreaseMAEMAE Increase
6h24h (original)0.0219–0.1031–
6h12h0.02357.31%0.10764.36%
6h6h0.026018.72%0.113710.28%
24h72h (original)0.0294–0.1197–
24h36h0.03084.76%0.12191.84%
24h18h0.03177.82%0.12363.26%
168h72h (original)0.0450–0.1539–
168h36h0.04694.22%0.15641.62%
168h18h0.04806.67%0.15842.92%
AuthorsOpenReview ↗

Rebuttal by Authors (Part 2 of 4)

Performance degrades consistently as the context is shortened, most at the 6-hour horizon. This reflects the strong short-term correlation of wind power: short-term forecasts rely more on recent observations, whereas long-term forecasts rely on the site's stable weather-to-power response, which a shortened context can still capture. Notably, even with one-quarter of the context, W2P-FM still substantially outperforms all full-context zero-shot baselines in Table 2 (e.g., MSE of 0.0260 vs. 0.0434 at 6h).

(2) Ablation of Meteo-Power Encoding. Extending the analysis in Figure 5(b), we retrain W2P-FM without MPE on the full pretraining corpus while retaining static context conditioning, using the same sites, partitions, configurations, and three random seeds.

SettingHorizonMSE (W2P-FM)MSE (w/o MPE)Increase (%)MAE (W2P-FM)MAE (w/o MPE)Increase (%)
Full-shot6h0.01890.01932.120.09280.09350.75
Full-shot24h0.02500.02552.000.10720.10861.31
Full-shot168h0.04410.04757.710.14860.15393.57
Zero-shot6h0.02190.02305.020.10310.10623.01
Zero-shot24h0.02940.03157.140.11970.12524.59
Zero-shot168h0.04500.052516.670.15390.16738.71

Removing MPE consistently degrades performance, especially in cross-site zero-shot forecasting, indicating that unseen sites rely more on MPE to capture site-specific weather-to-power responses from historical observations. The gap also generally widens with horizon, consistent with the analysis above: long-term forecasts depend more on the stable response patterns that MPE summarizes from the full context, complementing the local aggregation of MSFFN.

W2 & Q2: Anomalous Power Readings and Short-Gap Interpolation

Regarding the preprocessing concerns in W2 and Q2, we have revised Appendix C.1 to clarify anomaly identification and interpolation boundaries. Interpolation and sample construction never cross the training, validation, or test boundaries, and short-gap interpolation has a negligible impact on the overall evaluation results.

  • Invalid power observations, including missing or empty values (NaN), negative power readings, and periods explicitly marked as invalid operating conditions in the original records, are replaced with zeros.
  • The complete power series of each site is first divided chronologically into training, validation, and test sets, and short-gap interpolation and sliding-window construction are then performed independently within each split. Consequently, interpolation and sample construction never cross the training, validation, or test boundaries.
  • Since short-gap interpolation uses neighboring valid observations (Appendix C.1), one edge case may still arise within individual samples: an interpolated power value near the end of the historical context may be computed using a valid observation after the forecast origin. Such cases are rare: only about 0.02%, 0.005%, and 0.0007% of forecast target timestamps for the 6h, 24h, and 168h horizons, respectively, have their observed values used in such interpolation.
  • Using the original W2P-FM and W2P-FM-NS checkpoints from Table 1 without retraining, we reset interpolated power values to zero only in the historical context (History) or only in the forecast targets (Targets), while retaining the same test samples and forecast target timestamps. Removing short-gap interpolation leads to only marginal changes in MAE and MSE, and the relative performance of W2P-FM and W2P-FM-NS remains unchanged across all forecasting horizons.
SettingHorizonModelMSE (Original)MSE (History)MSE (Targets)MAE (Original)MAE (History)MAE (Targets)
Full-shot6hW2P-FM-NS0.02100.02100.02120.09860.09870.0991
Full-shot6hW2P-FM0.01890.01890.01910.09280.09280.0933
Full-shot24hW2P-FM-NS0.02790.02790.02820.11440.11430.1148
Full-shot24hW2P-FM0.02500.02500.02520.10720.10720.1076
Full-shot168hW2P-FM-NS0.04690.04690.04720.15410.15410.1546
Full-shot168hW2P-FM0.04410.04410.04430.14860.14860.1491
Zero-shot6hW2P-FM-NS0.02280.02280.02290.10540.10540.1057
Zero-shot6hW2P-FM0.02190.02190.02200.10310.10320.1034
Zero-shot24hW2P-FM-NS0.02980.02980.03000.12130.12130.1216
Zero-shot24hW2P-FM0.02940.02940.02950.11970.11970.1199
Zero-shot168hW2P-FM-NS0.04540.04540.04560.15510.15510.1555
Zero-shot168hW2P-FM0.04500.04500.04520.15390.15390.1542
AuthorsOpenReview ↗

Rebuttal by Authors (Part 3 of 4)

W3 & Q3: Seed-Level Variation and Paired Site-Level Analysis

Following your suggestion, we report seed-level standard deviations and conduct paired site-level analyses with bootstrap confidence intervals, and find that the improvements of W2P-FM are consistent across seeds and sites: it outperforms the best-performing baseline on more than 81% of full-shot sites and 97.30%–100% of unseen sites, and all 95% confidence intervals of the mean improvement remain above zero. We have added the paired site-level analysis and bootstrap confidence intervals to Appendix F (Table 19).

(1) Seed-level variation. Seed-level standard deviations of all full-shot methods are reported in Appendix F (Tables 13–15). In the North American 24-hour setting you highlighted, W2P-FM achieves an MSE of 0.0277 ± 0.0004 vs. 0.0282 ± 0.0001 for the strongest baseline (LightGBM), with a 3.81% MAE improvement (0.1137 vs. 0.1182). The three-seed mean and standard deviation of W2P-FM and W2P-FM-NS are summarized below; zero-shot baselines use official pretrained checkpoints without additional training and thus have no seed variation.

SettingHorizonW2P-FM (MSE)W2P-FM (MAE)W2P-FM-NS (MSE)W2P-FM-NS (MAE)
Full-shot6h0.0189±0.00000.0928±0.00030.0210±0.00010.0986±0.0003
Full-shot24h0.0250±0.00010.1072±0.00020.0279±0.00010.1144±0.0004
Full-shot168h0.0441±0.00020.1486±0.00030.0469±0.00020.1541±0.0000
Zero-shot6h0.0219±0.00010.1031±0.00020.0228±0.00000.1054±0.0000
Zero-shot24h0.0294±0.00030.1197±0.00080.0298±0.00000.1213±0.0005
Zero-shot168h0.0450±0.00040.1539±0.00070.0454±0.00030.1551±0.0008

(2) Paired site-level analysis. For each site and horizon, W2P-FM is compared against the best-performing baseline for that site and horizon. We report the site-level win rate and the mean and median relative improvements averaged over three random seeds. 95% confidence intervals of the mean improvement are estimated by paired bootstrap resampling over sites with 10,000 iterations.

SettingHorizonMetricWin RateMean Imp.Median Imp.95% CI
Full-shot6hMAE779/938 (83.05%)7.29%8.04%[5.90%, 8.37%]
Full-shot6hMSE787/938 (83.90%)7.57%8.81%[4.75%, 11.01%]
Full-shot24hMAE790/938 (84.22%)6.37%7.31%[4.94%, 7.54%]
Full-shot24hMSE766/938 (81.66%)5.17%6.39%[3.58%, 7.63%]
Full-shot168hMAE826/938 (88.06%)8.10%8.02%[7.11%, 9.05%]
Full-shot168hMSE817/938 (87.10%)9.55%10.76%[7.73%, 11.24%]
Zero-shot6hMAE74/74 (100%)26.78%25.49%[24.68%, 28.83%]
Zero-shot6hMSE74/74 (100%)44.74%43.98%[41.70%, 47.68%]
Zero-shot24hMAE73/74 (98.65%)31.08%28.76%[28.00%, 34.14%]
Zero-shot24hMSE73/74 (98.65%)47.85%43.88%[43.46%, 52.14%]
Zero-shot168hMAE72/74 (97.30%)28.64%27.37%[26.31%, 30.88%]
Zero-shot168hMSE73/74 (98.65%)43.72%42.25%[40.57%, 46.74%]
Reviewer_5n76OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces W2P-FM, a domain-specific foundation model for wind-power forecasting that learns shared weather-to-power representations across sites, forecast horizons, and temporal resolutions. Its architecture combines adaptive fusion of multiple numerical weather prediction sources, granularity-aware patching, static site-context conditioning, and historical weather–power response encoding, followed by variable-wise attention and multiscale temporal convolutions. Task-specific input/output components support deterministic, probabilistic, and fine-resolution forecasting.

Validation uses 1,012 wind-power series from 33 datasets across five continents, including full-shot forecasting, held-out-site and held-out-region zero-shot evaluation, and frozen-backbone adaptation to different temporal resolutions and probabilistic outputs. The authors report improvements over task-specific and pretrained forecasting baselines, supported by component ablations, model and training-data scaling analyses, and inference-time comparisons.

Critical strengths
  1. Broad and relevant empirical coverage. The corpus contains 1,012 series from 33 datasets across five continents. Evaluating both held-out sites and entirely held-out regional datasets is substantially more informative than testing only future intervals at training sites.

  2. A coherent architecture motivated by the application. Aligning patches by physical duration, incorporating future weather forecasts, and conditioning on recent site responses are sensible choices. The design addresses identifiable sources of heterogeneity rather than simply increasing model size.

  3. Useful evidence of representation transfer. Tables 13–14 show that adapting only the probabilistic head or the input/output projections can outperform competing methods. This is among the strongest findings because it directly supports the usefulness of the pretrained backbone.

  4. Meaningful ablations and reporting. The appendix includes uniform weather-source averaging, temporal-module ablations, model-size and data-volume experiments, and three-run variability for several main comparisons. It also acknowledges negative transfer from static context and limitations of representing regional generation through a single coordinate.

Critical weaknesses
  1. The comparisons confound architecture, input information, and training scope. W2P-FM is trained jointly across sites and horizons, whereas most trainable baselines are trained at the dataset level, and LightGBM is trained separately per site. W2P-FM also uses multiple weather sources, while baselines generally receive the single source selected on validation data. These are legitimate comparisons between complete forecasting systems, but they do not establish how much improvement comes from the proposed learning method. W2P-FM-NS removes static information but retains the advantages of multi-source weather inputs and pooled training. A strong pooled baseline trained on the same corpus with equivalent weather inputs, alongside a single-source W2P-FM comparison.

  2. Forecast-origin information availability is insufficiently documented. The paper identifies ERA5 as the historical meteorological source and GFS, EC, and Tianji as future forecast sources, but does not provide sufficient detail about initialization times, publication delays, historical forecast archives, or alignment between forecast issuance and prediction origin. This matters because a retrospective experiment must distinguish information available at the forecasting time from information obtained later. The dates of satellite imagery and annual AlphaEarth embeddings also require clarification relative to the evaluated periods.

  3. Preprocessing changes the evaluation population and may alter genuine zero-generation events.

Missing and anomalous observations are replaced with zeros; short zero runs are interpolated; windows are excluded when zeros exceed one third of either the historical or target interval. As written, this procedure does not clearly distinguish genuine zero output from missing or erroneous measurements. It also selects evaluation windows using the future target interval. Consequently, the reported accuracy describes a filtered population whose relationship to operational forecasting is unclear. Applying the same filtering to all models preserves a common comparison set, but does not resolve representativeness. Separate validity masks, retention rates by dataset, clarification of whether genuine zeros are interpolated, and confirmation that interpolation and window splitting cannot introduce cross-boundary information.

  1. The methodological novelty is primarily in integration, and its distinctive benefit needs stronger isolation. The model combines attention-based source weighting, resolution-specific projections, gated static conditioning, attention pooling, and dilated temporal convolutions. The combination is well motivated, but the manuscript does not yet establish a sufficiently clear general learning insight beyond the integrated system. The uniform-averaging ablation is valuable: Table 8 attributes approximately 3.7%–4.0% average MSE improvement to learned routing. However, this also demonstrates why the much larger headline improvements should not be attributed to routing alone. .

  2. The transfer claims need more precise interpretation. Cross-site “zero-shot” forecasting still uses 24 or 72 hours of target-site power history. That is a valid zero-shot protocol with respect to parameter updates, but it does not establish forecasting at sites with no operational history. Likewise, cross-granularity adaptation learns resolution-specific input/output components. It demonstrates efficient supervised adaptation, rather than direct generalization to arbitrary unseen resolutions. The paper should explicitly identify which adaptation sites and time intervals were already represented in hourly pretraining.

Questions for authors
  1. Were all weather forecasts issued and available before the prediction origin? How were initialization times and publication delays handled?

  2. Are genuine zero-power observations distinguishable from missing values, and what fraction of test windows was excluded?

  3. Were chronological splits formed before window construction and interpolation, with no training targets overlapping validation or test targets?

  4. What results are available under matched training corpora and weather-source access?

  5. How were future covariates supplied to each foundation-model baseline, and which checkpoints were used?

  6. Which adaptation sites and periods overlap with pretraining, and how were cross-region validation data selected?

9 comments
AuthorsOpenReview ↗

Rebuttal by Authors (Part 1 of 3)

Dear Reviewer 5n76,

We sincerely appreciate your thorough review and constructive feedback. We are glad that you highlight the broad and relevant empirical coverage, the coherent architecture motivated by the application, and the useful evidence of representation transfer.

W1 & Q4: Isolating the Architectural Contribution under Matched Training Data and NWP Inputs

Following your suggestion, we retrain three representative baselines on the same multi-site corpus as W2P-FM-NS with identical NWP inputs, and find that W2P-FM-NS still achieves the lowest MAE and MSE across all three horizons, reducing average errors by 5.76%–6.56% relative to the strongest baseline. These results demonstrate a clear advantage of its weather-to-power representation learning architecture. Moreover, comparing W2P-FM-NS under different NWP inputs shows that MeteoRouter effectively exploits the complementary information in multi-source NWP. The related experiments and analyses have been added to Appendix C.13.

(1) Controlled comparison

  • Data: All models use the same 938 pretraining sites, chronological data partitions, and training samples; static site context is excluded from all models.
  • NWP inputs: GFS only, or Uniform (the equally weighted average of the available NWP sources).
  • Training: Crossformer, PatchTST-GRU, and W2P-FM-NS are jointly trained across the context–horizon configurations (24, 6), (72, 24), and (72, 168); LightGBM, which does not support variable-length configurations, is trained separately for each horizon. Results (MAE/MSE) are averaged over three random seeds.
NWP inputHorizonW2P-FM-NSCrossformerPatchTST-GRULightGBM
Uniform6h0.0999/0.02180.1070/0.02350.1098/0.02430.1143/0.0251
Uniform24h0.1156/0.02870.1255/0.03080.1307/0.03250.1390/0.0348
Uniform168h0.1564/0.04750.1649/0.05040.1657/0.05010.1753/0.0530
GFS6h0.1040/0.02300.1104/0.02430.1107/0.02470.1173/0.0269
GFS24h0.1229/0.03030.1306/0.03330.1343/0.03420.1420/0.0376
GFS168h0.1568/0.04790.1684/0.05050.1661/0.05040.1762/0.0537

(2) W2P-FM-NS under different NWP inputs

The full-shot MSE of W2P-FM-NS decreases consistently from GFS to Uniform input and further to MeteoRouter-based fusion (W2P-FM-NS in Table 1), indicating that multi-source NWP provides complementary information and that MeteoRouter's adaptive source weighting exploits it more effectively. This is consistent with Table 8, where MeteoRouter outperforms uniform averaging for the full W2P-FM with static context.

HorizonGFSUniformMeteoRouter
6h0.02300.02180.0210
24h0.03030.02870.0279
168h0.04790.04750.0469

W2 & Q1: Clarification of Forecast-Origin Information Availability

All meteorological inputs are restricted to information preceding each forecast origin, and static site features predate the test periods. Future inputs rely solely on NWP runs initialized before the origin, and historical meteorological inputs are used only up to the origin. We have supplemented Appendices C.2 and C.3 with these details.

  • Future meteorological inputs: Future inputs are the publicly available historical forecasts issued by GFS, ECMWF-IFS (EC), and Tianji (Appendix C.2). For each forecast origin, we use the most recent NWP run initialized before that origin, consistent with common practice in wind power forecasting studies (Yakoub et al., 2023; Zhao et al., 2025).
  • Historical meteorological inputs: As in prior work (Ding et al., 2024; Abad-Santjago et al., 2025), we use ERA5 reanalysis (Hersbach et al., 2020) as a unified historical source, since the local observations of our 33 datasets vary in quality, completeness, variable availability, and temporal resolution. ERA5 is used only within the historical context up to the forecast origin. ERA5 and NWP fields are temporally aligned with the historical and target power timestamps, respectively, and mapped to each site via bilinear interpolation (Hardy and Finney, 2025).
  • Static site features: For the vast majority of sites, we use the most recent satellite imagery and annual AlphaEarth embedding preceding the start of each site's power record. For a few sites whose records begin particularly early, the earliest available features are used, which still predate the corresponding test periods; since static site characteristics change slowly, they remain representative. Zero-valued matrices are used when no features are available.
AuthorsOpenReview ↗

Rebuttal by Authors (Part 2 of 3)

W3 & Q2: Handling of Zero-Power Observations and Evaluation Windows

(1) Zero handling and interpolation

  • Missing, empty, or NaN power observations, negative readings, and source-identified invalid operating periods are set to zero.
  • Only runs of at most two consecutive zeros are linearly interpolated using neighboring valid observations, similar to previous wind power forecasting studies (Yan et al., 2022; Xu et al., 2025a).
  • Although genuine zero-power observations are not explicitly separated from invalid ones, longer zero runs remain unchanged, so sustained zero-output periods are preserved.

(2) Evaluation window filtering

  • A window is retained only if zeros account for no more than one third of its historical and target intervals, similar to the OpenWPF benchmark (Xu et al., 2025b), to ensure sufficient usable observations for model training and reliable error calculation.
  • This criterion is applied identically to all models and excludes only 9.22% (full-shot) and 10.62% (cross-site zero-shot) of candidate test windows, so about 90% of candidate windows are retained in the evaluation.

We have refined the description of these procedures in Appendix C.1.

Q3: Chronological Splitting and Interpolation Sensitivity

Each site's power series is first split chronologically into training, validation, and test partitions (7:1:2), and interpolation and sliding-window construction are then performed independently within each partition. This prevents cross-split information leakage and ensures non-overlapping forecast targets across the three sets. Additional sensitivity analyses further show that short-gap interpolation has a negligible impact on the overall evaluation results.

  • We note that one edge case may still arise within individual samples: an interpolated historical power value near the forecast origin may occasionally depend on an observation after that origin. Such cases are rare: only about 0.02%, 0.005%, and 0.0007% of forecast target timestamps for the 6h, 24h, and 168h horizons, respectively, have their observed values used in such interpolation.
  • To assess whether short-gap interpolation materially affects the reported forecasting results, we conduct two additional sensitivity analyses using the original W2P-FM and W2P-FM-NS checkpoints from Table 1 without retraining: interpolated values are reset to zero only in historical inputs (History) or only in forecast targets (Targets), while the same test samples and evaluated target timestamps are retained. Results are averaged over three random seeds. MAE and MSE show only marginal changes, and the relative performance of W2P-FM and W2P-FM-NS remains unchanged.
SettingHorizonModelMSE (Original)MSE (History)MSE (Targets)MAE (Original)MAE (History)MAE (Targets)
Full-shot6hW2P-FM-NS0.02100.02100.02120.09860.09870.0991
Full-shot6hW2P-FM0.01890.01890.01910.09280.09280.0933
Full-shot24hW2P-FM-NS0.02790.02790.02820.11440.11430.1148
Full-shot24hW2P-FM0.02500.02500.02520.10720.10720.1076
Full-shot168hW2P-FM-NS0.04690.04690.04720.15410.15410.1546
Full-shot168hW2P-FM0.04410.04410.04430.14860.14860.1491
Zero-shot6hW2P-FM-NS0.02280.02280.02290.10540.10540.1057
Zero-shot6hW2P-FM0.02190.02190.02200.10310.10320.1034
Zero-shot24hW2P-FM-NS0.02980.02980.03000.12130.12130.1216
Zero-shot24hW2P-FM0.02940.02940.02950.11970.11970.1199
Zero-shot168hW2P-FM-NS0.04540.04540.04560.15510.15510.1555
Zero-shot168hW2P-FM0.04500.04500.04520.15390.15390.1542
AuthorsOpenReview ↗

Rebuttal by Authors (Part 3 of 3)

W4: Core Design Insight and Validation of Key Components

We are encouraged that you find the combination well motivated.

W2P-FM's core insight is that heterogeneous weather and power inputs can be mapped into a unified temporal-variable token grid and conditioned on site-specific information, so that transferable weather-to-power representations are learned and decoupled from specific tasks. Its components are designed around three sources of heterogeneity in wind power forecasting:

  • Multi-source NWP and multi-resolution observations: MeteoRouter adaptively performs variable-wise fusion of multiple NWP sources, and Granularity-Aware Patch Embedding aligns multi-resolution sequences by keeping the patch duration fixed.
  • Site-dependent power responses: Site Feature Tokenization and Adaptive Static Context Injection incorporate static site characteristics, and Meteo-Power Encoding distills the site's recent weather-to-power response patterns.
  • Diverse forecasting requirements: A shared Variable-wise Multi-scale Transformer backbone is jointly pretrained across horizons and transferred to new granularities and probabilistic outputs through lightweight adaptation.

The benefits of these designs are isolated as follows:

  • Weather-to-power learning architecture. In the controlled comparison of our response to W1 & Q4, where all models are trained on the same multi-site corpus with identical NWP inputs, W2P-FM-NS still reduces average MAE and MSE by 5.76%–6.56% relative to the strongest baseline. Its individual designs are further supported by dedicated ablations (Figure 5(b); Tables 3 and 12).
  • Multi-source NWP fusion and static context conditioning. MeteoRouter outperforms uniform averaging by 3.66%–3.99% in MSE and remains effective under varying NWP availability (Tables 4 and 8; Figure 4), while static context conditioning generally improves performance and enhances generalization to unseen sites (Tables 1–2; Figure 5(a)).

W5: Precise Interpretation of the Transfer Settings

Our zero-shot setting evaluates generalization to sites unseen during pretraining, using recent historical power observations as context, without updating model parameters on target sites. This setting targets newly commissioned sites that lack sufficient data for model training but can provide recent observations, and is consistent with existing zero-shot forecasting studies such as WindFM (Fan et al., 2025) and Chronos-2 (Ansari et al., 2025).

Our cross-granularity and probabilistic experiments evaluate lightweight supervised adaptation: the pretrained backbone remains frozen, and only the granularity-specific power patch projection and forecasting head, or a new quantile head, are trained on downstream data.

W5 & Q6: Adaptation Sites and Periods Represented in Pretraining

As suggested, we have reported which adaptation sites and periods were already represented in hourly pretraining (Appendix C.8).

  • Sites: 39 of 55 cross-granularity sites and 47 of 77 probabilistic sites were represented in hourly pretraining.
  • Periods: For previously seen sites, both pretraining and downstream adaptation use only the first 70% of each site's records, so validation and test periods are never used for parameter updates. Since the downstream baselines are trained on the same periods, this overlap gives W2P-FM no information unavailable to the baselines.

Q5: Future Covariates and Checkpoints of Foundation-Model Baselines

We have added these configuration details to Appendix C.6.

  • Future covariates: Chronos-2, TiRex-2, CITRAS-FM, and Moirai 2.0 receive future NWP inputs through their native future-covariate interfaces, whereas WindFM, which does not support future NWP inputs, is evaluated with its original input configuration.
  • Checkpoints: All foundation-model baselines use their default officially released pretrained checkpoints, namely the standard Chronos-2 (120M), TiRex-2, and WindFM models and the only publicly released weights of Moirai 2.0 and CITRAS-FM; in the downstream cross-granularity and probabilistic comparisons, Chronos-2 and Moirai 2.0 are further fine-tuned from these checkpoints.

Q6: Cross-Region Validation Data

We have added these details to Appendix C.7. For cross-region zero-shot evaluation, entire target datasets are held out, and a separate W2P-FM model is pretrained and selected using only the training and validation splits of the remaining datasets. For zero-shot baselines, the NWP source is selected on the validation splits of the target datasets, which favors the baselines.


We appreciate your rigorous review. With these additional experiments and clarifications, we hope your concerns have been addressed, and we would be grateful if you would consider updating your score. We welcome any further feedback.

Reviewer_vv7bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Weather W2P-FM, a foundation model for wind power forecasting. The model operates in two phases. In first phase, the model accumulates heterogeneous data sources (eg, nwp) to create a unified representation. This is then fed to a novel variable-wise transformer to predict power readings. Experimental validation is done against a range of baselines for 1000+ site-level wind power series worldwide.

Critical strengths
  • The paper considers a problem of important scientific relevance, and proposes a fully data-driven approach to tackling it. Experiments are done extensively comparing against a range of prior work to show the benefits of the proposed approach, and the results show reasonable improvement.
Critical weaknesses
  • The presentation is very unclear at many unplaces. For example, in section 4.1, 4.2, many domain specific steps for data fusion are produced without much motivation. Rather, more time is spent explaining every layer of the neural network, which is information that could easily be deferred to an appendix.
  • Ignores references to a large body of work in weather forecasting, including foundation model approaches such as Climax (Nguyen et al.) and Aurora. Even though they do not predict energy directly, they can be finetuned for this task directly.
Questions for authors

See weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer vv7b,

Thank you for your review and helpful suggestions. We are encouraged that you recognize the important scientific relevance of the problem and the extensive experiments against a range of prior work.

W1: Clarification of Methodological Motivation and Improved Presentation

All components of W2P-FM are motivated by a main goal: learning transferable weather-to-power representations across sites and tasks. This goal is hindered by three types of heterogeneity in wind power forecasting: (i) multi-source NWP and multi-resolution observations, (ii) site-dependent power responses, and (iii) diverse forecasting requirements. Accordingly, W2P-FM maps heterogeneous inputs into a unified temporal-variable token grid, conditions them on site-specific context, and decouples a shared backbone from task-specific forecast heads, as detailed below.

(1) Unified Representation of Heterogeneous Multi-Source Wind-Power Data

  • Challenge: NWP sources vary in availability and reliability across variables, regions, and lead times, while power and meteorological observations have different temporal resolutions (Lines 78–84).
  • Design: MeteoRouter adaptively performs variable-wise fusion of multiple NWP sources, and Granularity-Aware Patch Embedding aligns multi-resolution sequences by keeping the patch duration fixed (Section 4.1; validated in Section 5.3, Lines 431–457).

(2) Site-Conditioned Weather-to-Power Learning for Cross-Site Generalization

  • Challenge: Identical meteorological forcing can induce markedly different power responses across sites due to variations in terrain, capacity, and local environmental conditions (Lines 84–89).
  • Design: Site Feature Tokenization and Adaptive Static Context Injection incorporate static site characteristics, and Meteo-Power Encoding distills the site's recent weather-to-power response patterns (Section 4.2; analyzed in Section 5.3, Lines 458–476).

(3) Flexible Task Adaptation Across Diverse Forecasting Requirements

  • Challenge: Forecasting applications differ in horizons, temporal granularities, and output formulations (Lines 89–93).
  • Design: A shared Variable-wise Multi-scale Transformer backbone, jointly pretrained across horizons, is decoupled from task-specific forecast heads (Sections 4.3–4.4 and 5.1), improving accuracy across horizons and transferring to new granularities and probabilistic objectives (Sections 5.2–5.3).

Following your suggestion, we have refined the Introduction and Sections 4.1–4.2 to make the motivation and core mechanism of each component more prominent, and moved the lower-level details of Granularity-Aware Patch Embedding and Adaptive Static Context Injection to Appendix D.

W2: Relationship to Weather Foundation Models and Their Integration as Upstream Inputs

Weather foundation models such as ClimaX (Nguyen et al., 2023) and Aurora (Bodnar et al., 2025) represent important advances in data-driven weather forecasting. They were not included in our baseline comparison because they differ fundamentally from W2P-FM in task formulation and in their role within the forecasting pipeline, and cannot be readily adapted to site-level wind power forecasting through fine-tuning or simple variants. Specifically:

  • ClimaX and Aurora learn representations of the atmospheric state from gridded fields spanning many variables and pressure levels and predict its future evolution, whereas W2P-FM targets site-level wind power generation, learning transferable weather-to-power representations from historical power observations, historical meteorological data, future NWP trajectories, and static site context (Section 3, Lines 142–157). The two thus differ in formulation, inputs, and prediction targets.
  • Within the forecasting pipeline, weather foundation models act as upstream components that produce meteorological forecasts, which can in turn be used as inputs to wind power forecasting models such as W2P-FM.
  • Direct fine-tuning is also of limited suitability, as their architectures take gridded atmospheric fields as inputs and do not natively accept historical power observations, which are needed to capture weather-to-power responses (Lines 84–89).

In future work, weather foundation models can be naturally integrated into W2P-FM as upstream inputs. Since the projection in MeteoRouter is shared across sources, their forecasts can be incorporated as additional sources, or used in place of conventional NWP products, without architectural changes. We have added this discussion to Appendix C.2, and plan to systematically evaluate this integration in subsequent studies.


Thank you again for your review. We hope the clarified design motivation, the improved presentation, and the discussion of weather foundation models address your concerns, and would be grateful if you would reconsider your score. We are happy to discuss any remaining questions.

AuthorsOpenReview ↗

Gentle Reminder: Looking Forward to Your Feedback

Dear Reviewer vv7b,

Thank you again for your time and helpful suggestions. As the discussion period is drawing to a close, we would like to follow up on our response. Following your comments, we have:

  • Revised the Introduction and Sections 4.1–4.2 to emphasize the motivation for each component, and moved the lower-level details to Appendix D.
  • Added a discussion of weather foundation models such as ClimaX and Aurora, clarifying their role as upstream forecasters and how they can be integrated into W2P-FM through MeteoRouter (Appendix C.2).

We would greatly appreciate it if you could take a look at the revised manuscript and let us know whether these changes address your concerns. If any points remain unclear, we would be happy to provide further clarification before the discussion closes. Should you find our responses satisfactory, we would be grateful if you would consider reassessing your evaluation.

Best regards,

The Authors

Comments on the paper

2 comments
Area_Chair_4VrqOpenReview ↗

Check the Response

Dear Reviewers,

The authors have submitted their response to your review comments. Please kindly take a moment to read their response and assess whether your concerns have been adequately addressed. Further discussion with the authors to reach a deeper mutual understanding would be greatly appreciated.

Thanks.

AuthorsOpenReview ↗

Summary of Rebuttal and Discussion

Dear Area Chair and Reviewers,

We sincerely thank you for your careful reviews and constructive feedback. Our paper makes three main contributions:

  • We introduce W2P-FM, a foundation model for wind power forecasting that learns transferable weather-to-power representations across sites and tasks.
  • Its well-motivated architecture combines availability-aware multi-source NWP fusion and granularity-aware tokenization with static context conditioning and historical response encoding. Furthermore, a Variable-wise Multi-scale Transformer is designed to integrate intra-patch cross-variable interactions with local multi-scale inter-patch aggregation.
  • Extensive experiments on 1,012 site-level series from 33 datasets across five continents demonstrate strong forecasting performance, robust zero-shot generalization to unseen sites and regions, and efficient adaptation to downstream tasks.

During the rebuttal, we added controlled comparisons under matched training data and NWP inputs, sensitivity analyses of historical context length and short-gap interpolation, MPE ablation results, and statistical evidence of robustness across seeds and sites. We also clarified information availability and preprocessing, baseline configurations, and the protocols and scope of zero-shot evaluation and downstream adaptation. The revisions further explain the design motivation and the relationship to weather foundation models.

Reviewers 5n76 and 3EP8 acknowledged that our clarifications and additional experiments helped address their concerns, with both updating their ratings to weak acceptance. We have not yet received further feedback from Reviewer vv7b following our response to their comments on presentation and related work and the corresponding revisions to the manuscript.

The constructive feedback throughout the discussion has helped us refine the manuscript and improve its rigor and clarity. Thank you again for your time and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.