Pith. sign in

REVIEW 3 major objections 7 minor 114 references

Cross-farm wind power curves transfer best when source turbines are selected by distributional similarity, not geographic distance.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:49 UTC pith:6DT4LP5Z

load-bearing objection A sensible source-selection idea, clearly written, but the evaluation leaks target labels into the weights and K-choice; the headline margins need a nested, source-only redo before they can be trusted. the 3 major comments →

arxiv 2607.19744 v1 pith:6DT4LP5Z submitted 2026-07-22 stat.AP cs.LG

Domain-Adapted Power Curve for Cross-Farm Applications

Ahmadreza Chokhachian , V. Roshan Joseph , Yu Ding This is my paper
classification stat.AP cs.LG
keywords domain adaptationpower curvetransfer learningsource selectiondissimilarity metricwind farm planningaccumulated local effectsSCADA data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a domain-adaptation method for building power curves on a new wind farm that has no power measurements yet. The central claim is that selecting source turbines on an existing farm via a weighted dissimilarity metric—combining distributional differences in environmental covariates and terrain descriptors, weighted by their importance for power output—and ensembling their individual power-curve models gives substantially lower prediction error than pooling all turbines or choosing sources by proximity. Evaluated with leave-one-turbine-out cross-validation and a distant-farm planning split on a 66-turbine farm, the method consistently beats terrain-aware Gaussian processes, graph neural networks, binning, and geographic-neighbor transfer. If correct, it implies that site-planning power curves should be built from turbines experiencing similar inflow conditions rather than from the nearest or most topographically similar farm.

Core claim

The paper's central discovery is that cross-farm power curve transfer is best achieved through supervised distributional source matching. Each turbine's domain is defined by its temporal environmental covariates (wind speed, temperature, wind direction, turbulence intensity, direction standard deviation) and static terrain descriptors (slope, RIX, ridge height). The weighted dissimilarity metric d_WD compares these domains using Kolmogorov-Smirnov distances for covariates and absolute differences for terrain features, with weights derived from accumulated local effects (ALE) of each feature on power output. Selecting the K=7 source turbines with the smallest d_WD and averaging the prediction

What carries the argument

The key machinery is the weighted dissimilarity metric d_WD (Eq. 8). For each terrain feature it uses the absolute difference of min-max scaled values; for each environmental covariate it uses the Kolmogorov-Smirnov distance between empirical marginal distributions. These per-feature distances are combined with ALE-based feature weights (wind speed 77.7%, RIX 9.5%, temperature 5.4%, ridge 5.2%, slope 2.1%). This metric identifies the K most similar source turbines, and predictions are aggregated by averaging the outputs of per-source power-curve models (the ensemble strategy). The metric is computationally light at O(n log n) per pair, about 4200 times faster than Sinkhorn Wasserstein distan

Load-bearing premise

The load-bearing assumption is that the ALE feature weights used in the dissimilarity metric can be computed without the target farm's power observations; in the paper's experiments those weights are derived from the full dataset including the held-out target turbine, so the reported advantage may shrink if weights are recomputed from source data alone.

What would settle it

Retrain the ALE weights for each leave-one-turbine-out fold using only the 65 training turbines' data (excluding the target's power values), then measure the LOTO RMSE of the WD-based ensemble; if the advantage over geographic-neighbor transfer collapses or reverses, the metric's success depends on test-label leakage.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Site-planning power curves should be transferred from turbines with similar wind and terrain distributions, not from geographically close farms.
  • The ALE-based weighting automatically discounts irrelevant features, so the method adapts to different farms where terrain matters more.
  • The O(n log n) dissimilarity makes the approach scalable to large SCADA archives and many candidate source farms.
  • The ensemble-of-selected-sources strategy is a generally applicable recipe for transfer learning when source heterogeneity is present.
  • The method provides a principled alternative to optimal transport and graph-based transfer for regression models in other domains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The evaluation may overstate the benefit: ALE weights are computed from the full dataset, including held-out target turbines' power readings; in a real planning scenario those readings are unavailable, so recomputing weights from source-only data is needed to verify the advantage.
  • The method is validated on a single wind farm with an artificial 'distant farm' split; its performance on truly separate farms with tens of kilometers of separation remains untested.
  • Given that wind speed dominates the weights, a simplified metric using wind speed and RIX alone might recover much of the gain and is worth testing as a cheaper baseline.
  • The paper assumes same make, model, and service duration across turbines; the transfer mechanism may not extend to different turbine types without additional adaptation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes a domain-adaptation approach for transferring wind-turbine power curves from an operating farm to a planning farm. The method defines a weighted dissimilarity metric d_WD that combines Kolmogorov–Smirnov distances for temporal environmental covariates and absolute differences for static terrain features, with feature weights derived from accumulated local effects (ALE). The K most similar source turbines are selected and their individual power-curve models are ensembled to predict at the target. The method is evaluated on a single 66-turbine farm using leave-one-turbine-out cross-validation and a 'distant farm planning' (DFP) split, and is compared against pooled, geographic-neighbor, and literature baselines. The authors report consistent improvements and claim the approach is suitable for site-planning power predictions.

Significance. If the reported gains are real, the paper would offer a principled, computationally cheap alternative to geographic-distance-based transfer, with a clear mechanism for source selection. The manuscript provides public code and data, and compares against a wide range of baselines, which are strengths. However, the current empirical evidence does not support the cross-farm claim as stated: the evaluation appears to use held-out power labels when computing the ALE weights, and K is selected on the same LOTO folds that are later reported as the method's performance. These are internal-validity threats that could substantially inflate the advantage of WD transfer. The single-farm design also limits external validity. The central idea remains plausible, but the experiments need to be redone with a source-only, nested protocol before the central claim can be accepted.

major comments (3)
  1. [Section 3.1, Table 1, Eqs. (8)–(10)] The ALE feature weights w_c are obtained from a model g fitted to predict power output. The manuscript reports a single global weight table and does not state that w_c is recomputed inside each LOTO or DFP fold using only the training turbines. If the global weights are used, the held-out turbine's own y-values influence the dissimilarity metric d_WD that selects that turbine's sources; in the DFP experiment the seven target turbines' power labels contribute to the same global weights. In the real site-planning scenario the target farm has no power observations, so the evaluated pipeline differs from the deployed pipeline. This can bias the comparison in favor of WD transfer. Please re-run all experiments with weights estimated from source data only, or otherwise demonstrate that the fixed weights are independent of the test turbines, and report whether the rankings in Tables 3 and 4 per
  2. [Section 2, Section 3.2, Figure 5, Table 3] The number of source turbines K=7 is chosen by minimizing the average LOTO RMSE over K (Figure 5), and the same averaged LOTO RMSE is then presented in Table 3 as the proposed method's performance. This makes the headline RMSE an in-sample selection statistic rather than an out-of-sample estimate. The DFP experiment is equally affected because K is selected using all 66 turbines, including the seven DFP target turbines. A nested evaluation is needed: select K on a validation split (or via an inner CV) and report RMSE on an untouched test set. At minimum, report the sensitivity of the conclusions to the K-selection protocol.
  3. [Abstract, Section 4.2, Conclusion] The experiments are conducted entirely within a single wind farm. The DFP split uses a geographically separated cluster of the same farm (~2 km away) as a proxy for a planning farm, and the authors acknowledge this limitation. While the proxy experiment is useful, the abstract and title claim 'cross-farm applications' and 'consistently outperforms' for site-planning. As it stands, the empirical evidence supports only within-farm transfer among turbines of the same make and model, with no test on an independent farm. Please either temper the claims to match the evidence or add an independent-farm validation if possible.
minor comments (7)
  1. [Section 3.1] Please specify the training data and hyperparameters for the ALE model g (e.g., which turbines/years were used, covariance specification, optimization details). This is needed for reproducibility.
  2. [Eq. (8) and Table 1] Eq. (8) sums weights to 1, while Table 1 reports weights summing to 100. Clarify the normalization used, or use consistent scaling.
  3. [Section 4.1, text after Table 3] The reported percentage reductions from pooling to WD transfer appear too large. For ANN: (3.86−3.30)/3.86 ≈ 14.5% (2017) and (4.03−3.50)/4.03 ≈ 13.2% (2018), not 15–17%. Similar discrepancies occur for thinned twinGP and XGBoost. Please recompute and correct the percentages.
  4. [Supplementary Material] The supplementary section says the data and code are on a GitHub page but does not provide a URL or repository identifier. Please add the actual link.
  5. [Table 1] Column header 'W eight' has a typo; it should be 'Weight'.
  6. [Figure 5] The average LOTO RMSE curves are point estimates without variability. Consider adding error bars or confidence bands, especially since the differences between K values are small.
  7. [Section 4.1] The GNN baseline is listed as an own implementation because the original code is not public. Please note any potential differences from Daenens et al. (2025) that could affect the comparison.

Circularity Check

2 steps flagged

Reported WD-transfer gains are partly fitted on target labels: global ALE weights and K=7 are chosen on the same LOTO folds later reported as performance.

specific steps
  1. fitted input called prediction [Section 3.1, Eq. (8) and Eq. (10), Table 1; LOTO evaluation in Section 4.1, Table 3]
    "Computing ALE ranges requires a fitted model to serve as g ... The fitted model uses wind speed, temperature, slope, RIX, and ridge height to predict power output. Table 1 reports the resulting ALE main effects for each feature. ... d_WD[(X^(i),s_i),(X^(target),s_target)] = sum_{c in C_s} w_c d^(s)_c(i,target) + sum_{c in C_x} w_c d^(x)_c(i,target)"

    Eq. (10) defines the ALE importance rho_c from a model g fitted to predict power y, and Eq. (8) uses the resulting weights w_c in the dissimilarity metric that selects source turbines. The paper reports a single global weight table (Table 1) and does not state that weights are recomputed inside each LOTO fold using only training turbines. Thus the held-out target turbine's own y-values can influence the metric that selects its own sources. Since the paper's central claim is site-planning prediction where the target has no power observations, this makes the reported transfer advantage partly fitted on test labels rather than obtainable from source data alone.

  2. fitted input called prediction [Section 2 (LOTO as design guide), Section 3.2 (Figure 5, K selection), Table 3 (headline LOTO results)]
    "By averaging over all turbines, LOTO provides a stable error estimate and is therefore well-suited for guiding design choices such as the aggregation strategy and the number of source turbines. ... RMSE decreases steadily as K grows from 2 to 7. Beyond that, improvement stalls and becomes inconsistent, with occasional small increases. K=7 is a practical sweet spot."

    The same average LOTO RMSE that is used in Figure 5 to select K=7 is later reported in Table 3 as the out-of-sample performance of 'The proposed WD Transfer (K=7)'. This means the headline RMSE and the reported advantage over baselines are in-sample selection outcomes, not independent test-set estimates. At minimum, a nested procedure—selecting K inside each LOTO fold—is needed before the claimed superiority can be treated as a genuine prediction rather than a fit to the evaluation criterion.

full rationale

The proposed weighted-dissimilarity metric itself is not circular by construction: the KS distances in Eq. (7) compare covariate distributions and are well-defined without target power labels, and the ensemble prediction in Eq. (13) is a legitimate transfer procedure. However, the paper's empirical support for the central claim has two evaluation-level reductions. First, the ALE feature weights entering Eq. (8) are derived from a model fitted to power output; as reported, one global weight table is used, so in the LOTO experiment the held-out turbine's y can affect the source-selection metric. In the real site-planning scenario no target y exists, so either weights must be estimated from source data alone or the reported gains reflect test-label leakage. Second, K=7 is chosen by minimizing the same LOTO RMSE that is later presented as the method's performance in Table 3, making the headline comparison optimistically biased. These issues do not make the method definitionally circular, and the paper includes independent baselines and public-code reproducibility, so a score of 6 rather than higher is appropriate. The self-citations to Chokhachian et al. (2026a,b) are used as computational tools and baselines, not as uniqueness arguments, and are not load-bearing in a circular way beyond the leakage described.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. Its free parameters are the source-count K, the ALE-derived feature weights, and downstream model hyperparameters. The main unstated assumptions are homogeneity of turbines, sufficiency of marginal distributional matching, validity of ALE importance weights, and the proxy status of the DFP split.

free parameters (3)
  • K (number of source turbines) = 7
    Selected in Section 3.2 by minimizing average LOTO RMSE in Figure 5 on the same data later used for final RMSE tables; this is a data-fitted parameter and the reported test errors are not nested-cross-validated.
  • ALE feature weights w_c = wind speed 77.7%, temperature 5.4%, slope 2.1%, RIX 9.5%, ridge 5.2% (Table 1)
    Estimated from all 66 turbines via an STGP model; used in Eq. (8). Because the estimation uses the same data as the LOTO evaluation, it can leak target information.
  • Downstream model hyperparameters (thinned twinGP, ANN, XGBoost, thinned SV) = not reported
    No hyperparameter settings are given; performance comparisons depend on these choices.
axioms (5)
  • domain assumption Turbines on the farm are homogeneous in make, model, and service duration, so a power curve learned from one turbine is applicable to another.
    Assumed in Section 2 and used throughout; if false, cross-turbine power-curve transfer is confounded by hardware differences.
  • domain assumption Marginal distributional similarity of each covariate (KS distance per variable) captures the domain shift relevant for power-curve transfer.
    Eq. (7) uses marginal KS distances only; joint dependencies (e.g., wind speed × direction) are not explicitly matched.
  • domain assumption ALE main-effect range is a valid measure of covariate importance for weighting the dissimilarity metric.
    Section 3.1 uses ALE ranges (Eq. 9-10) to set w_c; this assumes the black-box model g is accurate enough for its local derivatives to reflect predictive importance.
  • domain assumption A 2.2 km-isolated cluster within one wind farm (DFP) is a valid proxy for a planning farm with no operational data.
    Section 4.2; the authors explicitly admit the separation is not that of two real wind farms.
  • domain assumption The ALE weight model and selection metric are valid when computed without target power labels.
    In the real site-planning scenario the target has no power output, but the paper's global weights may have been computed using the target's labels in LOTO; the paper does not show source-only weights work as well.

pith-pipeline@v1.3.0-alltime-deepseek · 14099 in / 13112 out tokens · 127135 ms · 2026-08-01T11:49:28.571283+00:00 · methodology

0 comments
read the original abstract

The wind energy industry relies on accurate power curve models to make power forecast, evaluate turbine performance, quantify upgrade, or support site-planning decisions. In this paper, we focus on site-planning power curves, i.e., we investigate how power curve models trained using turbine data on an operating wind farm can be transferred to a new, undeveloped farm. The traditional wisdom in the wind energy literature relies on distance, layout, or terrain characteristics for making cross-farm power curve transfer. Through the lens of domain adaptation, we propose a more reliable transfer learning approach for cross-farm power curve modeling. In the cross-farm applications, a domain is specified by the temporal environmental variates and spatial terrain variables. Domain adaptation is to find a capable similarity metric to adapt the domain on the new farm to that on the existing farm. Empirical results show that our domain adapted power curve consistently outperforms competing approaches by an appreciable margin for site-planning power predictions.

Figures

Figures reproduced from arXiv: 2607.19744 by Ahmadreza Chokhachian, V. Roshan Joseph, Yu Ding.

Figure 1
Figure 1. Figure 1: Conceptual illustration of turbine-specific power curves (left panel) and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Wind farm layout The original data were recorded at a higher frequency and then averaged to 10-minute intervals to align with the standard SCADA (Supervisory Control and Data Acquisition) resolution. For each year, a complete dataset would contain 52,560 observations per turbine. However, after preprocessing, approximately 40,000 to 45,000 observations remain for each turbine, depending on the extent of mi… view at source ↗
Figure 3
Figure 3. Figure 3: Concatenation strategy. Data from all selected source turbines are pooled into a [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ensemble strategy. A separate model is fitted for each selected source turbine, [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Average LOTO RMSE over different K’s and either aggregation strategy. The left two plots use the thinned twinGP, whereas the right two plots use a feedforward neural network. In each panel the orange curve is the ensemble strategy and the blue curve is the concatenation strategy. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: For target turbines # 38 (in the left plot) and # 39 (in the right plot), the seven [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

114 extracted references · 1 linked inside Pith

  1. [1]

    Roshan Joseph and Yu Ding , title =

    Ahmadreza Chokhachian and V. Roshan Joseph and Yu Ding , title =. Technometrics , volume =

  2. [2]

    Energy , volume=

    A novel data-driven deep learning approach for wind turbine power curve modeling , author=. Energy , volume=

  3. [3]

    Wiley Interdisciplinary Reviews: Computational statistics , volume=

    Energy distance , author=. Wiley Interdisciplinary Reviews: Computational statistics , volume=

  4. [4]

    Advances in neural information processing systems , volume=

    Faster Wasserstein distance estimation with the Sinkhorn divergence , author=. Advances in neural information processing systems , volume=

  5. [5]

    Kusiak and H

    A. Kusiak and H. Zheng and Z. Song. On-line monitoring of power curves. Renewable Energy. 2009

    2009

  6. [6]

    R. K. Pandit AND A. J. Kolios , journal=

  7. [7]

    Energy Conversion and Management , volume=

    A new approach to very short term wind speed prediction using k-nearest neighbor classification , author=. Energy Conversion and Management , volume=

  8. [8]

    IEEE Transactions on Sustainable Energy , volume=

    Time adaptive conditional kernel density estimation for wind power forecasting , author=. IEEE Transactions on Sustainable Energy , volume=. 2012 , publisher=

    2012

  9. [9]

    Frontiers in Energy Research, section Wind Energy , volume =

    Ding, Yu and Barber, Sarah and Hammer, Florian , title =. Frontiers in Energy Research, section Wind Energy , volume =

  10. [10]

    and Tibshirani, R

    Hastie, T. and Tibshirani, R. and Friedman, J. , title =

  11. [11]

    Chipman, H. A. and George, E. I. and McCulloch, R. E. , title =. The Annals of Applied Statistics , volume =

  12. [12]

    , title =

    Vapnik, V. , title =

  13. [13]

    and Ding, Y

    Lee, G. and Ding, Y. and Xie, L. and Genton, Marc G. , title =. Wind Energy , volume =

  14. [14]

    and Galloway, S

    Stephen, B. and Galloway, S. J. and McMillan, D. and Hill, D. C. and Infield, D. G. , title =. IEEE Transactions on Power Systems , volume =

  15. [15]

    and Parthasarathy, G

    Uluyol, O. and Parthasarathy, G. and Foslien, W. and Kim, K. , title =. Proceedings of the Annual Conference of the Prognostics and Health Management Society , volume =

  16. [16]

    and Ding, Y

    Lee, G. and Ding, Y. and Genton, Marc G. and Xie, L. , title =. Journal of the American Statistical Association , volume =. 2015 , doi =

    2015

  17. [17]

    Proceedings of the 24th International Conference on Machine Learning , pages =

    Dai, Wenyuan and Yang, Qiang and Xue, Gui-Rong and Yu, Yong , title =. Proceedings of the 24th International Conference on Machine Learning , pages =

  18. [18]

    Ahmadreza Chokhachian and Matthias Katzfuss and Yu Ding , journal =. Fast. 2026 , volume =

    2026

  19. [19]

    Evaluation of wind power generation potential using a three hybrid approach for households in

    Mojtaba Qolipour and Ali Mostafaeipour and Shahaboddin Shamshirband and Omid Alavi and Hossein Goudarzi and Dalibor Petković , journal =. Evaluation of wind power generation potential using a three hybrid approach for households in

  20. [20]

    2022 , volume =

    Hammer, Florian and Barber, Sarah , title =. 2022 , volume =

    2022

  21. [21]

    475--480

    HOMER: The Hybrid Optimization Model for Electric Renewables , author =. 1995 , pages = "475--480", note =

    1995

  22. [22]

    2022 , author =

    Statistical-analytical study on world development trend in offshore wind energy production capacity focusing on Great Britain with the aim of MCDA based offshore wind farm siting , journal =. 2022 , author =

    2022

  23. [23]

    2017 , author =

    Application of the GIS-DANP-MABAC multi-criteria model for selecting the location of wind farms: A case study of Vojvodina, Serbia , journal =. 2017 , author =

    2017

  24. [24]

    2018 , author =

    A multi-criteria GIS based model for wind farm site selection using interval type-2 fuzzy analytic hierarchy process: The case study of Nigeria , journal =. 2018 , author =

    2018

  25. [25]

    Transfer strategy for power output estimation of wind farm at planning stage based on a

    Li, Zihao and Sun, Wei and Xiang, Yue and Harrison, Gareth P , journal=. Transfer strategy for power output estimation of wind farm at planning stage based on a. 2022 , publisher=

    2022

  26. [26]

    2025 , volume =

    Daenens, Simon and Verstraeten, Timothy and Daems, Pieter-Jan and Nowé, Ann and Helsen, Jan , title =. 2025 , volume =

    2025

  27. [27]

    Renewable Energy , volume =

    Spatially transferable machine learning wind power prediction models: v. Renewable Energy , volume =. 2024 , author =

    2024

  28. [28]

    2019 , author =

    A novel method for wind farm layout optimization based on wind turbine selection , journal =. 2019 , author =

    2019

  29. [29]

    2018 , author =

    A probabilistic assessment approach for wind turbine-site matching , journal =. 2018 , author =

    2018

  30. [30]

    Energies , volume=

    Applications and modeling techniques of wind turbine power curve for wind farms—A review , author=. Energies , volume=

  31. [31]

    , TITLE =

    Vahidzadeh, Mohsen and Markfort, Corey D. , TITLE =. Energies , VOLUME =. 2020 , NUMBER =

    2020

  32. [32]

    2025 , note =

    Clerc, Alex and Calvo, Gabriele , title =. 2025 , note =

    2025

  33. [33]

    Barber and Y

    S. Barber and Y. Ding , title =. Proceedings of the 2024 WindEurope Annual Event, Bilbao, Spain, March 20-22 , year =

    2024

  34. [34]

    2025 , author =

    Inter-turbine modelling of wind-farm power using multi-task learning , journal =. 2025 , author =

    2025

  35. [35]

    IISE Transactions , volume=

    When wind travels through turbines: A new statistical approach for characterizing heterogeneous wake effects in multi-turbine wind farms , author=. IISE Transactions , volume=. 2017 , publisher=

    2017

  36. [36]

    IEEE Transactions on power systems , volume=

    Direction-dependent power curve modeling for multiple interacting wind turbines , author=. IEEE Transactions on power systems , volume=. 2017 , publisher=

    2017

  37. [37]

    IEEE Transactions on Artificial Intelligence , volume=

    A decade survey of transfer learning (2010--2020) , author=. IEEE Transactions on Artificial Intelligence , volume=

    2010

  38. [38]

    Journal of Machine Learning Research , volume=

    Domain-Adversarial Training of Neural Networks , author=. Journal of Machine Learning Research , volume=

  39. [39]

    ECCV Workshops , pages=

    Deep CORAL: Correlation Alignment for Deep Domain Adaptation , author=. ECCV Workshops , pages=

  40. [40]

    Advances in neural information processing systems , volume=

    Adversarial multiple source domain adaptation , author=. Advances in neural information processing systems , volume=

  41. [41]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep cocktail network: Multi-source unsupervised domain adaptation with category shift , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  42. [42]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Moment matching for multi-source domain adaptation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  43. [43]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Multi-source distilling domain adaptation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  44. [44]

    ICML , pages=

    Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptation , author=. ICML , pages=

  45. [45]

    arXiv:2112.06745 , year=

    A Survey of Unsupervised Domain Adaptation for Visual Recognition , author=. arXiv:2112.06745 , year=

    Pith/arXiv arXiv

  46. [46]

    IEEE TPAMI , year=

    A Comprehensive Survey on Source-Free Domain Adaptation , author=. IEEE TPAMI , year=

  47. [47]

    SIAM/ASA Journal on Uncertainty Quantification , volume =

    Katzfuss, Matthias and Guinness, Joseph and Lawrence, Earl , title =. SIAM/ASA Journal on Uncertainty Quantification , volume =

  48. [48]

    and Georgiopoulos, M

    Bebis, G. and Georgiopoulos, M. , journal=. Feed-forward neural networks , year=

  49. [49]

    Roshan Joseph , title =

    Chaofan Huang and V. Roshan Joseph , title =. Technometrics , volume =

  50. [50]

    Energy Statistics

    Sz. Energy Statistics. A Class of Statistics Based on Distances , journal =. 2013 , volume =

    2013

  51. [51]

    and Rasch, Malte J

    Gretton, Arthur and Borgwardt, Karsten M. and Rasch, Malte J. and Sch. A Kernel Two Sample Test , journal =. 2012 , volume =

    2012

  52. [52]

    , title =

    Rubner, Yossi and Tomasi, Carlo and Guibas, Leonidas J. , title =. International Journal of Computer Vision , year =

  53. [53]

    European conference on computer vision , pages=

    Deep coral: Correlation alignment for deep domain adaptation , author=. European conference on computer vision , pages=

  54. [54]

    Computational Optimal Transport , publisher =

    Peyr. Computational Optimal Transport , publisher =. 2019 , volume =

    2019

  55. [55]

    2012 , doi =

    Sugiyama, Masashi and Suzuki, Taiji and Kanamori, Takafumi , title =. 2012 , doi =

    2012

  56. [56]

    Journal of Machine Learning Research , year =

    Kanamori, Takafumi and Hido, Shohei and Sugiyama, Masashi , title =. Journal of Machine Learning Research , year =

  57. [57]

    Machine Learning , year =

    Breiman, Leo , title =. Machine Learning , year =

  58. [58]

    , title =

    Wolpert, David H. , title =. Neural Networks , year =

  59. [59]

    Renewables 2024: Analysis and forecast to 2030 , year =

    2024

  60. [60]

    electricity generation by energy source , year =

    What is U.S. electricity generation by energy source , year =

  61. [61]

    Electricity generation from wind , year =

  62. [62]

    2024 , institution =

    Wiser, Ryan and Bolinger, Mark and others , title =. 2024 , institution =

    2024

  63. [63]

    Part 12-1: Power performance measurements of electricity producing wind turbines , organization =

    IEC 61400-12-1: Wind energy generation systems. Part 12-1: Power performance measurements of electricity producing wind turbines , organization =. 2022 , url =

    2022

  64. [64]

    Wind Energy , year =

    Astolfi, Davide and Castellani, Francesco and Lombardi, Lorenzo and Marri, Luigi , title =. Wind Energy , year =

  65. [65]

    Technometrics , year =

    Prakash, Abhinav and Tuo, Rui and Ding, Yu , title =. Technometrics , year =

  66. [66]

    and Ding, Yu , title =

    Prakash, Abhinav and Lee, Se Yoon and Liu, Xin and Liu, Ling and Mallick, Bani K. and Ding, Yu , title =. IEEE Transactions on Sustainable Energy , year =. doi:10.1109/TSTE.2023.3328374 , note =

    arXiv 2023

  67. [67]

    IEEE Transactions on Knowledge and Data Engineering , year =

    Pan, Sinno Jialin and Yang, Qiang , title =. IEEE Transactions on Knowledge and Data Engineering , year =

  68. [68]

    and Wang, DingDing , title =

    Weiss, Karl and Khoshgoftaar, Taghi M. and Wang, DingDing , title =. Journal of Big Data , year =

  69. [69]

    , title =

    Massey, Frank J. , title =. Journal of the American Statistical Association , year =

  70. [70]

    Muandet, Krikamol and Fukumizu, Kenji and Sriperumbudur, Bharath K. and Sch. Kernel Mean Embedding of Distributions: A Review and Beyond , journal =. 2017 , volume =

    2017

  71. [71]

    Zhou, Zhi-Hua , title =

  72. [72]

    Ding and J

    Y. Ding and J. Tang and J. Z. Huang , title =. Proceedings of ASME Turbo Expo 2015: Turbine Technical Conference and Exposition (GT 2015), Montreal, Canada, June 15-19 , year =

    2015

  73. [73]

    Journal of Statistical Software , year =

    Zammit-Mangion, Andrew and Cressie, Noel , title =. Journal of Statistical Software , year =

  74. [74]

    Fixed rank

    Cressie, Noel and Johannesson, Gardar , journal=. Fixed rank

  75. [75]

    2000 , pages =

    Olea, Ricardo , title =. 2000 , pages =

    2000

  76. [76]

    Journal of Statistical Planning and Inference , volume=

    Energy statistics: A class of statistics based on distances , author=. Journal of Statistical Planning and Inference , volume=

  77. [77]

    A. M. Sempreviva and S. E. Larsen and N. G. Mortensen and I. Troen , title =. Boundary-Layer Meteorology , volume =

  78. [78]

    Han and J

    X. Han and J. Guo and P. Wang , title =. IET Generation, Transmission and Distribution , volume =

  79. [79]

    Tian and A

    W. Tian and A. Ozbay and H. Hu , title =. Procedia Engineering , volume =

  80. [80]

    Mallick , title =

    Yu Ding and Abhinav Prakash and Se Yoon Lee and Xin Liu and Lei Liu and Bani K. Mallick , title =. Renewable Energy , volume =

Showing first 80 references.