Pith. sign in

REVIEW 5 major objections 8 minor 37 references

Three-Stage Composite Outlier Identification of Wind Power Data: Integrating Physical Rules with Regression Learning and Mathematical Morphology

T0 review · 5 major / 8 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Wind-power data outliers can be found by a three-stage pipeline combining physical rules, RANSAC regression, and image morphology.

desk verdict A plausible three-stage composition of known outlier-identification tools, with a clean write-up and some independent labeled checks, but the headline metric can reward over-removal and needs label-based confirmation. read the letter →

arxiv 2504.21354 v1 pith:BSXE4N33 submitted 2025-04-30 eess.SP

Limengqian Zheng , Lipeng Zhu , Weijia Wen , Jiayong Li , Cong Zhang This is my paper · ORCID
classification eess.SP
keywords outlieridentificationwindpowercurvephysicalrulesRANSACmathematicalmorphologySCADAdatapredictionstackedoutliers
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that no single outlier detector is enough for wind-turbine data and that a deliberate sequence of three complementary passes can clean the data reliably: physical rules first remove impossible points, RANSAC-based regression with adaptive thresholds removes most scattered and stacked outliers, and mathematical morphology on an image of the data catches the dense outlier clusters that remain. If the method works as claimed, wind-farm operators could clean SCADA-style wind speed and power measurements with little manual threshold tuning. The paper reports that the cleaned data yield wind-power curve fitting errors no larger than 0.037 per unit and an F1 score above 0.94 on a labeled dataset, with downstream wind-power prediction errors reduced by about 20%.

What carries the argument

The load-bearing mechanism is the image-morphology envelope check. The remaining wind speed-power points are normalized, quantized into a $(q+1) imes(q+1)$ binary matrix, and treated as a digital image; a circular structuring element of diameter $d=6$ is used to erode the image and then dilate the eroded result, removing thin noise while preserving the main band of normal data. The upper and lower boundaries of the nonzero pixels in each column are interpolated into a smooth envelope, and any original point outside the envelope is flagged as an outlier. This stage is what the paper credits with catching dense stacked outliers that the regression stage misses.

What would settle it

Construct a synthetic dataset whose normal points form two separated bands (for example, two power setpoints at overlapping wind speeds), run the morphology stage alone, and count how many injected normal points fall outside the interpolated envelope; a false-positive rate much higher than on a single-band control dataset would show that the single-envelope assumption is load-bearing.

Watch

Extended reading notes

Core claim

The paper's central claim is that a three-stage composite outlier identification method can reliably and adaptively identify both dispersive outliers and stacked outliers in wind speed-power data, including stacked outliers densely packed near normal data. Stage one applies simple physical operating rules, such as rejecting negative wind speed, negative power, nonzero power below cut-in speed, and zero power above cut-in speed. Stage two trains a cubic-polynomial RANSAC regression model on stratified samples and uses interquartile-range thresholds on prediction errors to flag most remaining anomalies. Stage three rasterizes the residual data into a binary image, applies erosion and dilation with a circular structuring element, interpolates the upper and lower boundary of the surviving nonzero pixels, and marks every original point outside that envelope as an outlier. Numerical tests on three real turbines and one simulated dataset are used to argue that the composite method beats LOF, Quartile-DBSCAN, MDUE, regression-learning alone, and morphology alone, while retaining computational efficiency and improving wind-power prediction accuracy.

Load-bearing premise

The final stage assumes that after shrinking and regrowing the point cloud, every legitimate wind-speed-power point lies inside one continuous band; if legitimate data form separate clusters or thin tails, the method will flag valid points as outliers.

Editorial extensions

If this is right

  • Wind-farm operators can clean operational data without hand-tuning detection thresholds: the IQR step derives its thresholds from the data, and the morphology step adapts to the shape of the normal power curve.
  • Cleaned wind-power curves on the tested turbines keep normalized root-mean-square fitting error at or below 0.037 per unit, compared with 0.058 to 0.141 for the comparison methods on the same datasets.
  • On the labeled turbine dataset, accuracy stays above 96% and F1 above 0.94; on the simulated dataset, accuracy stays above 94% and F1 reaches 0.928.
  • Removing and interpolating the identified outliers reduces short-term wind-power prediction MAE and RMSE by roughly 20% relative to using the unprocessed dataset.
  • The paper argues the same three-stage scheme can be extended to other types of wind-turbine SCADA measurements beyond wind speed and power.

Reading between the lines

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

  • A natural extension would reuse the morphology stage as a generic bivariate data-quality filter wherever normal operation forms a band, such as rotor speed versus power or temperature versus output.
  • The method is only partly 'adaptive': the IQR factor $k$ and the structuring-element diameter $d$ still come from user-chosen default ranges, so a fully data-driven selection of these two parameters would be the next step.
  • Because the cleaned points retain the natural vertical scatter of the power curve instead of collapsing to a single line, the output could also feed condition-monitoring tasks that look for deviations from a normal operating band.
  • The strongest confirmation of the dense-stacked-outlier claim would come from a controlled benchmark with injected outlier fractions at known densities, rather than only the one labeled dataset used here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper proposes a three-stage composite outlier identification method for wind speed-power SCADA data. Stage 1 applies physical rules to remove points violating basic turbine operating principles (negative speed/power, zero power above cut-in, nonzero power below cut-in, missing/duplicate records). Stage 2 fits a cubic polynomial wind speed-power model using RANSAC with stratified sampling and 5-fold cross-validation, then flags points whose prediction errors fall outside IQR-based adaptive thresholds. Stage 3 converts the remaining data into a binary image, applies erosion and dilation with a circular structuring element, extracts an envelope from the cleaned image, and marks points outside the envelope as outliers. The method is evaluated on five real SCADA datasets (WT1-WT5) and one simulated dataset (WT6), with comparisons to LOF, Quartile-DBSCAN, MDUE, regression learning alone, and mathematical morphology alone. Metrics include wind power curve fitting RMSE, classification accuracy/F1 on the labeled WT3 dataset, wind power prediction error after cleaning, and computation time. The authors conclude that the method is reliable, adaptive, and superior to the baselines, with F1 above 0.94 on labeled data, fitting errors below 0.037 p.u., and a 20% improvement in prediction accuracy.

Significance. The problem addressed is practically important: SCADA wind power data are heavily contaminated, and existing outlier identification methods often fail on stacked outliers near normal data. The proposed integration of physical rules, robust regression, and morphological image processing is a sensible and reasonably novel combination, and the paper provides a substantial experimental study across multiple real and simulated datasets. The labeled WT3 dataset and the simulated WT6 dataset with known ground truth are valuable assets for verification. The method is computationally efficient enough for practical use. However, the significance of the claimed superiority is currently undermined by evaluation design issues: the primary comparative metric on unlabeled datasets rewards over-removal, the regression model is structurally misspecified at the plateaus of the power curve, and the morphological envelope assumption may discard legitimate data in sparse or multimodal distributions. If these concerns are addressed with additional label-based evaluations and robustness analyses, the method could be a useful contribution to wind power data quality improvement.

major comments (5)
  1. [Section IV-B, Table III; Eq. (22)] The central comparative metric on unlabeled datasets WT1 and WT2 is the wind power curve fitting RMSE (Ermse in Eq. (22)) computed after outlier removal. This metric structurally rewards any method that deletes points far from a smooth curve, regardless of whether those points are true outliers, and Stage 2 of the proposed method removes exactly the points with large regression residuals. Therefore the observed Ermse reduction is partly a consequence of the method's construction, and the claimed superiority over LOF, Quartile-DBSCAN, and MDUE on these datasets is not established. The authors should report label-based precision, recall, and F1 on datasets where ground truth is available or can be constructed (e.g., expert annotation on a subset of WT1/WT2, or additional synthetic labeled datasets), and should also report the fraction of points removed by each method so that over-removal can be assessed.
  2. [Section III-B, Eq. (7) and Eqs. (11)-(13)] The cubic polynomial regression model is structurally misspecified for the wind speed-power curve, which has a flat zero-power region below cut-in and a rated-power plateau at the top. After physical rule preprocessing, low-speed zero-power points and rated-power points remain in the data, and a single cubic polynomial cannot represent these plateaus. The residuals of legitimate points at the plateaus will be systematically large, and the IQR thresholds in Eqs. (11)-(13) will consequently flag valid boundary points as outliers. The authors should either replace the cubic model with a piecewise or physically constrained model, or demonstrate via labeled data that this misspecification does not cause a material number of false positives.
  3. [Section III-C, Steps 2 and 3; Eqs. (20)-(21)] The mathematical morphology refinement assumes that the normal data form a single connected envelope, so that erosion followed by dilation with a circular structuring element of diameter d=6 yields a band that contains all legitimate points. This assumption will fail if normal data are sparse, split into multiple separated clusters, or contain legitimate narrow tails in the wind speed-power distribution, and in those cases the method will mark valid data as outliers. The paper provides no analysis of how the envelope extraction depends on data density or cluster structure. The authors should test the method on simulated data with controlled normal distributions (varying density, multiple clusters, and narrow tails) and report precision/recall breakdowns to characterize the failure modes of this stage.
  4. [Section V; Section IV-A] The conclusion states that more than 30% of the points in the case studies are anomalous, but this contamination level is not discussed in the data description or in the threshold-setting stage. When more than 30% of the error distribution is contaminated, the quartiles Q1 and Q3 in Eq. (11) are themselves influenced by outliers, so the 'adaptive' IQR thresholds in Eqs. (12)-(13) are not guaranteed to provide reliable separation. The authors should quantify the contamination ratio for each dataset, analyze the robustness of the IQR thresholding as a function of contamination level, and justify why the thresholds remain appropriate under the observed heavy contamination.
  5. [Section IV-B, Table III and Fig. 13] All comparative performance metrics are reported as single point estimates without error bars, confidence intervals, or statistical significance tests. The RL stage of the proposed method uses 5-fold cross-validation, and the paper says the results are averaged over folds, yet only one value is presented. For a fair and convincing comparison, the authors should report mean and standard deviation over cross-validation folds or multiple runs for the proposed method and for stochastic baselines, and perform a significance test (e.g., paired t-test or Wilcoxon) for the differences in Table III and Fig. 13.
minor comments (8)
  1. [Table VI] The row 'WT5 100% 0% 1.000' appears twice; one duplicate should be removed.
  2. [Fig. 15] The x-axis tick labels appear corrupted as '43 5 6 7 8' and should read '3 4 5 6 7 8'.
  3. [Section III-B, Eq. (8)] The function Me(x) is described as 'estimates the mean value of all the elements in x', but Eq. (8) computes the mean absolute deviation around the median. The definition should be clarified.
  4. [Section III-C, Eqs. (17)-(19)] The floor function definition in Eq. (19) is nonstandard for negative values; please clarify with a concrete example how negative normalized values are mapped to integer coordinates.
  5. [References] References [2] and [24] are the same paper (F. Shojaei et al.) and should be merged or differentiated.
  6. [Table VII] For WT5, the proposed method reduces Ermse from 0.103 to 0.090, which is much smaller than the reduction seen on WT4 (0.130 to 0.027). The authors should explain this limited improvement on WT5, as it may indicate reduced effectiveness on certain data distributions.
  7. [Section IV-A] The statement that more than 30% of data points are anomalous should be introduced in the data description section rather than appearing only in the conclusion, so that the reader can interpret the subsequent results in that context.
  8. [Section III-D, Eq. (22)] The description of the bin method and cubic spline interpolation used to compute the power curve for Eq. (22) is brief; please provide enough detail to make the computation reproducible.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity in the unlabeled evaluation: post-cleaning curve-fit RMSE largely re-measures the method's own removal rule, though labeled F1 results are independent.

  1. fitted input called prediction [Section III-B Eq. (14) as evaluated by Section III-D Eq. (22) and Table III]
    "Ua ={(vi,pi)| (vi,pi)∈ U,ei≤tlow or ei≥tup} Un ={(vi,pi)| (vi,pi)∈ U,tlow <e i <t up} (14); Ermse = 1/C sqrt(1/N sum (ˆpi−pi)^2) (22). First, since the wind power curve fitting error is widely used ... it is chosen as the straightforward metric to initially assess the potential of the proposed methodology in removing outliers far from the normal wind power curve."

    Stage 2 defines the 'normal' set as points whose errors relative to the RANSAC-fitted cubic regression fall inside IQR thresholds. The main unlabeled evaluation then measures Ermse by fitting a bin-method power curve to the cleaned dataset. Because the cleaning rule explicitly discards points far from a smooth regression, the post-cleaning bin-curve fit is close to that regression by construction; the reported Ermse reduction is partly a built-in consequence of the removal rule rather than an independent test of whether those points are truly anomalous. On unlabeled WT1 and WT2, where no labels exist, comparing methods via Ermse rewards aggressive removal, so the claimed superiority over LOF, Quartile-DBSCAN, and MDUE is not fully established by those rows.

full rationale

The three-stage method itself is not definitionally circular: each stage operates on input data with stated rules, and no fitted parameter is simultaneously used as its own ground truth. The labeled WT3 and WT6 experiments give external F1/accuracy evidence that the central outlier-identification claim has independent content. However, the paper's unlabeled WT1/WT2 comparisons rest on post-cleaning wind-power-curve RMSE, and that metric is confounded with the cleaning rule: Eq. (14) removes points with large regression residuals, and Eq. (22) then measures how close the remaining points are to a smooth curve. Since the method is explicitly designed to keep points near a smooth curve, its low Ermse is partly guaranteed by construction, and methods that retain legitimate scatter are penalized. This is a real partial circularity in the evaluation logic, but it does not collapse the whole derivation, because the labeled metrics are independent and the method's stages do not import their conclusions from the paper's own citations. Overall score reflects partial evaluation circularity, not a fully self-referential derivation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. It relies on six hand-set numerical parameters, of which only k and d receive sensitivity analysis, and on several domain assumptions about the shape and separability of normal wind speed-power data. These assumptions are the main thing a re-implementation would need to test on turbines with different power curves, data densities, or curtailment behavior.

free parameters (6)
  • RANSAC threshold multiplier c = 1.43
    Set once on the basis of the authors' empirical tests and not subjected to sensitivity analysis in the paper.
  • RANSAC maximum iterations Nmax = 1000
    Chosen as a default typical value with no sensitivity study; affects model selection stability.
  • IQR adjustment factor k = 1.5
    Standard IQR multiplier; sensitivity is shown only on WT3, with a recommended range of 1.0 to 2.5.
  • Structuring element diameter d = 6
    Controls erosion and dilation in the morphology stage; sensitivity is shown only on WT3, with a recommended range of 4 to 7.
  • Image quantization level q = 100
    Sets the image resolution for the morphology stage; described as a typical value and not varied.
  • Polynomial expansion degree = 3
    Cubic polynomial chosen without comparison to other degrees or regularization; affects the Stage 2 regression shape.
assumptions (5)
  • domain assumption Normal wind speed-power data form a connected, band-like two-dimensional distribution after quantization.
    The MM refinement extracts upper and lower boundaries of non-zero pixels per column and marks anything outside the envelope as an outlier; if normal data are sparse, disconnected, or multimodal, the envelope becomes ill-defined. See Section III-C Steps 2 and 3.
  • domain assumption A single cubic polynomial can represent the normal wind speed-power relationship across the full operating range.
    Stage 2 trains a cubic RANSAC regression, but real power curves have a cut-in region, a steep rise, and a rated-power plateau, so a global cubic is a modeling choice. See Section III-B, Eqs. (5) to (10).
  • domain assumption Zero power above cut-in and nonzero power below cut-in are always anomalous.
    Physical rules (3) and (4) treat curtailment or stopped-turbine data above cut-in as outliers, even though the paper itself notes such data may be correct measurements under abnormal operating conditions. See Section III-A.
  • domain assumption The synthetic labels of WT3 are representative of real outlier distributions.
    The labeled dataset is made by manually injecting outliers into a completely normal dataset cleansed by domain experts; performance on it may not transfer to naturally occurring outliers. See Section IV-A.
  • standard math Standard properties of mathematical morphology erosion and dilation transfer to binary images of SCADA data.
    The method applies textbook erosion and dilation (Eqs. 20 and 21) and assumes these operations clean noise without destroying the normal envelope; this is standard but unverified for this specific data geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Three-Stage Composite Outlier Identification of Wind Power Data: Integrating Physical Rules with Regression Learning and Mathematical Morphology." pith.science (2026). https://pith.science/paper/BSXE4N33

@misc{pith2026250421354,
  author       = {Pith},
  title        = {Pith review of: Three-Stage Composite Outlier Identification of Wind Power Data: Integrating Physical Rules with Regression Learning and Mathematical Morphology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BSXE4N33}},
  note         = {Machine review of arXiv:2504.21354}
}
read the original abstract

Existing studies on identifying outliers in wind speed-power datasets are often challenged by the complicated and irregular distributions of outliers, especially those being densely stacked yet staying close to normal data. This could degrade their identification reliability and robustness in practice. To address this defect, this paper develops a three-stage composite outlier identification method by systematically integrating three complementary techniques, i.e., physical rule-based preprocessing, regression learning-enabled detection, and mathematical morphology-based refinement. Firstly, the raw wind speed-power data are preprocessed via a set of simple yet efficient physical rules to filter out some outliers obviously going against the physical operating laws of practical wind turbines. Secondly, a robust wind speed-power regression learning model is built upon the random sample consensus algorithm. This model is able to reliably detect most outliers with the help of an adaptive threshold automatically set by the interquartile range method. Thirdly, by representing the wind speed-power data distribution with a two-dimensional image, mathematical morphology operations are applied to perform refined outlier identification from a data distribution perspective. This technique can identify outliers that are not effectively detected in the first two stages, including those densely stacked ones near normal data points. By integrating the above three techniques, the whole method is capable of identifying various types of outliers in a reliable and adaptive manner. Numerical test results with wind power datasets acquired from distinct wind turbines in practice and from simulation environments extensively demonstrate the superiority of the proposed method as well as its potential in enhancing wind power prediction.

Figures

Figures reproduced from arXiv: 2504.21354 by the authors.

Figure 1
Figure 1. Distribution characteristics of outliers in wind power data. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Overall process of RL-enabled outlier detection. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Overall framework of the proposed methodology. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Flowchart of the MM-based refinement. ν˜i = f(νiq) + 1 (17) p˜i = f(pi q) + 1 (18) f(x) = ( [x], x ≥ 0 [x] + 1, x < 0 (19) where q is a positive integer that determines the range of the resulting integer values, and f(x) is a floor function, expressed as [19], with [x]…
Figure 5
Figure 5. Figure 5: Wind power curve of each wind power dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 8
Figure 8. Figure 8: Outlier identification results for WT3. (a) Identification based on [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Comparison of wind power curves. (a) Actual normal wind power [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 7
Figure 7. Figure 7: Outlier identification results for WT2. (a) Identification based on [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 10
Figure 10. Figure 10: Outlier identification results by LOF. Wind speed (m/s) -10 -5 0 5 10 15 20 25 0 500 1500 2000 (a) WT1 Power (kW) 1000 Wind speed (m/s) (b) WT2 0 Power (kW) 1000 2000 3000 0 Wind speed (m/s) 5 10 15 20 (c) WT3 0 500 1500 2000 Power (kW) 1000 -10 -5 0 5 10 15 20 25 nor…
Figure 11
Figure 11. Figure 11: Outlier identification results by Quartile-DBSCAN. [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Outlier identification results by MDUE. 75% MDUE LOF MM Quartile- DBSCAN RL Proposed method 0% 4% 0.08 0.12 0.16 20% 0.6 0.7 0.8 0.9 1.0 Accuracy/p.u. Error rate/p.u. F1 score/p.u. 80% 85% 90% 95% 100% 8% 12% 16% [PITH_FULL_IMAGE:figures/full_fig_p009_12.png]
Figure 13
Figure 13. Figure 13: Identification metrics for various methods in WT3. [PITH_FULL_IMAGE:figures/full_fig_p009_13.png]
Figure 14
Figure 14. Figure 14: Outlier identification results for WT3 with the variation of the [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 15
Figure 15. Figure 15: Outlier identification results for WT3 with the variation of the [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 16
Figure 16. Figure 16: Outlier identification results for WT4. (a) Outlier annotation by [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]
Figure 17
Figure 17. Figure 17: Outlier identification results for WT5. (a) Outlier annotation by [PITH_FULL_IMAGE:figures/full_fig_p011_17.png]
Figure 18
Figure 18. Figure 18: Outlier identification results for WT6. (a) True distribution of [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

  1. [1]

    Hierarchical model predictive control strategy based on dynamic active power dispatch for wind power cluster integration,

    L. Ye, C. Zhang, Y . Tang, W. Zhong, Y . Zhao, P. Lu, B. Zhai, H. Lan, and Z. Li, “Hierarchical model predictive control strategy based on dynamic active power dispatch for wind power cluster integration,” IEEE Trans. Power Syst., vol. 34, no. 6, pp. 4617–4629, 2019

    work page 2019

  2. [3]

    Sequential data-driven automatic calibration of wind turbine fault information in smart grids,

    L. Zhu and Y . Song, “Sequential data-driven automatic calibration of wind turbine fault information in smart grids,” IEEE Internet Things J. , vol. 9, no. 11, pp. 8943–8952, 2022

    work page 2022

  3. [4]

    Time series data-driven online prognosis of wind turbine faults in presence of scada data loss,

    L. Zhu and X. Zhang, “Time series data-driven online prognosis of wind turbine faults in presence of scada data loss,” IEEE Trans. Sustain. Energy, vol. 12, no. 2, pp. 1289–1300, 2021

    work page 2021

  4. [5]

    Wind power curve modeling and wind power forecasting with inconsistent data,

    Y . Wang, Q. Hu, D. Srinivasan, and Z. Wang, “Wind power curve modeling and wind power forecasting with inconsistent data,” IEEE Trans. Sustain. Energy, vol. 10, no. 1, pp. 16–25, 2019

    work page 2019

  5. [7]

    Wind power forecasting based on a spatialtemporal graph convolution network with limited engineering knowledge,

    L. Yang, F. Tsung, K. Wang, and J. Zhou, “Wind power forecasting based on a spatialtemporal graph convolution network with limited engineering knowledge,” IEEE Trans. Instrum. Meas. , vol. 73, pp. 1–13, 2024

    work page 2024

  6. [8]

    Low-voltage ride-through remote testing method for offshore wind turbines,

    Z. Jiang and Y . Liu, “Low-voltage ride-through remote testing method for offshore wind turbines,” IEEE Trans. Instrum. Meas., vol. 69, no. 6, pp. 2905–2913, 2020

    work page 2020

  7. [9]

    Collection system topol- ogy for deep-sea offshore wind farms considering wind characteristics,

    Y . Fu, Y . Liu, L.-l. Huang, F. Ying, and F. Li, “Collection system topol- ogy for deep-sea offshore wind farms considering wind characteristics,” IEEE Trans. Energy Convers. , vol. 37, no. 1, pp. 631–642, 2022

    work page 2022

  8. [10]

    Feature selection for binary classification within functional genomics experiments via interquartile range and clustering,

    Z. Khan, M. Naeem, U. Khalil, D. M. Khan, S. Aldahmani, and M. Hamraz, “Feature selection for binary classification within functional genomics experiments via interquartile range and clustering,” IEEE Access, vol. 7, pp. 78 159–78 169, 2019. 13

    work page 2019

Show all 37 references
  1. [11]

    Comparing threshold selection methods for network anomaly detection,

    A. Komadina, M. Martini, S. Gro, and . Mihajlovi, “Comparing threshold selection methods for network anomaly detection,” IEEE Access, vol. 12, pp. 124 943–124 973, 2024

    2024

  2. [12]

    An adaptive identification method of abnormal data in wind and solar power stations,

    H. Wang, N. Zhang, E. Du, J. Yan, S. Han, N. Li, H. Li, and Y . Liu, “An adaptive identification method of abnormal data in wind and solar power stations,” Renewable Energy, vol. 208, pp. 76–93, 2023

    2023

  3. [13]

    Models for monitoring wind farm power,

    A. Kusiak, H. Zheng, and Z. Song, “Models for monitoring wind farm power,” Renewable Energy, vol. 34, no. 3, pp. 583–590, 2009

    2009

  4. [14]

    Modeling wind-turbine power curve: A data partitioning and mining approach,

    T. Ouyang, A. Kusiak, and Y . He, “Modeling wind-turbine power curve: A data partitioning and mining approach,” Renewable Energy, vol. 102, pp. 1–8, 2017

    2017

  5. [15]

    A combined algorithm for cleaning abnor- mal data of wind turbine power curve based on change point grouping algorithm and quartile algorithm,

    X. Shen, X. Fu, and C. Zhou, “A combined algorithm for cleaning abnor- mal data of wind turbine power curve based on change point grouping algorithm and quartile algorithm,” IEEE Trans. Sustain. Energy, vol. 10, no. 1, pp. 46–54, 2018

    2018

  6. [16]

    Raw wind data preprocessing: A data- mining approach,

    L. Zheng, W. Hu, and Y . Min, “Raw wind data preprocessing: A data- mining approach,” IEEE Trans. Sustain. Energy, vol. 6, no. 1, pp. 11–19, 2015

    2015

  7. [17]

    Data-driven correction approach to refine power curve of wind farm under wind curtailment,

    Y . Zhao, L. Ye, W. Wang, H. Sun, Y . Ju, and Y . Tang, “Data-driven correction approach to refine power curve of wind farm under wind curtailment,” IEEE Trans. Sustain. Energy , vol. 9, no. 1, pp. 95–105, 2018

    2018

  8. [18]

    An abnormal data processing method based on an ensemble algorithm for early warning of wind turbine failure,

    Q. Zhao, K. Bao, Z. Wei, Y . Han, and J. Wang, “An abnormal data processing method based on an ensemble algorithm for early warning of wind turbine failure,” IEEE Access, vol. 9, pp. 69 044–69 060, 2021

    2021

  9. [19]

    Copula-based model for wind turbine power curve outlier rejection,

    Y . Wang, D. G. Infield, B. Stephen, and S. J. Galloway, “Copula-based model for wind turbine power curve outlier rejection,” Wind Energy, vol. 17, no. 11, pp. 1677–1688, 2014

    2014

  10. [20]

    Image-based abnormal data detection and cleaning algorithm via wind power curve,

    H. Long, L. Sang, Z. Wu, and W. Gu, “Image-based abnormal data detection and cleaning algorithm via wind power curve,” IEEE Trans. Sustain. Energy, vol. 11, no. 2, pp. 938–946, 2019

    2019

  11. [21]

    A fast abnormal data cleaning algorithm for performance evaluation of wind turbine,

    Z. Wang, L. Wang, and C. Huang, “A fast abnormal data cleaning algorithm for performance evaluation of wind turbine,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–12, 2021

    2021

  12. [22]

    Wind power curve data cleaning algorithm via image thresholding,

    Y . Su, F. Chen, G. Liang, X. Wu, and Y . Gan, “Wind power curve data cleaning algorithm via image thresholding,” in 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO) , 2019, pp. 1198– 1203

    2019

  13. [23]

    Wind power curve data cleaning by image thresholding based on class uncertainty and shape dissimilarity,

    G. Liang, Y . Su, F. Chen, H. Long, Z. Song, and Y . Gan, “Wind power curve data cleaning by image thresholding based on class uncertainty and shape dissimilarity,” IEEE Trans. Sustain. Energy , vol. 12, no. 2, pp. 1383–1393, 2021

    2021

  14. [24]

    Filters optimized tuning for wind farms reactive power calculation,

    F. Shojaei, H. Samet, and T. Ghanbari, “Filters optimized tuning for wind farms reactive power calculation,” IEEE Trans. Instrum. Meas. , vol. 70, pp. 1–9, 2021

    2021

  15. [25]

    Deep belief network-based hammer- stein nonlinear system for wind power prediction,

    F. Li, M. Zhang, Y . Yu, and S. Li, “Deep belief network-based hammer- stein nonlinear system for wind power prediction,” IEEE Trans. Instrum. Meas., vol. 73, pp. 1–12, 2024

    2024

  16. [26]

    Robust calibration approach for robot base coordinate system based on ransac-svd,

    G. Wang, J. Wen, X. Gao, L. Yan, and I.-M. Chen, “Robust calibration approach for robot base coordinate system based on ransac-svd,” IEEE Trans. Instrum. Meas. , vol. 72, pp. 1–9, 2023

    2023

  17. [27]

    An image matching op- timization algorithm based on pixel shift clustering ransac,

    S. Ma, P. Guo, H. You, P. He, G. Li, and H. Li, “An image matching op- timization algorithm based on pixel shift clustering ransac,” Information Sciences, vol. 562, pp. 452–474, 2021

    2021

  18. [28]

    An enhanced mutated particle filter tech- nique for system state estimation and battery life prediction,

    M. Ahwiadi and W. Wang, “An enhanced mutated particle filter tech- nique for system state estimation and battery life prediction,” IEEE Trans. Instrum. Meas. , vol. 68, no. 3, pp. 923–935, 2019

    2019

  19. [29]

    Dependency analysis of accuracy estimates in k-fold cross validation,

    T.-T. Wong and N.-Y . Yang, “Dependency analysis of accuracy estimates in k-fold cross validation,” IEEE Trans. Knowl. Data Eng. , vol. 29, no. 11, pp. 2417–2427, 2017

    2017

  20. [30]

    Performance prediction us- ing high-order differential mathematical morphology gradient spectrum entropy and extreme learning machine,

    H. Zhao, H. Liu, J. Xu, and W. Deng, “Performance prediction us- ing high-order differential mathematical morphology gradient spectrum entropy and extreme learning machine,” IEEE Trans. Instrum. Meas. , vol. 69, no. 7, pp. 4165–4172, 2020

    2020

  21. [31]

    Multi-focus image fusion through gradient- based decision map construction and mathe- matical morphology,

    X. Bai, M. Liu, Z. Chen, P. Wang, and Y . Zhang, “Multi-focus image fusion through gradient- based decision map construction and mathe- matical morphology,” IEEE Access, vol. 4, pp. 4749–4760, 2016

    2016

  22. [32]

    Method for cleaning abnormal data of wind turbine power curve based on density clustering and boundary extraction,

    Z. Luo, C. Fang, C. Liu, and S. Liu, “Method for cleaning abnormal data of wind turbine power curve based on density clustering and boundary extraction,” IEEE Trans. Sustain. Energy, vol. 13, no. 2, pp. 1147–1159, 2022

    2022

  23. [33]

    A normal behavior model based on power curve and stacked regressions for condition monitoring of wind turbines,

    F. Bilendo, H. Badihi, N. Lu, P. Cambron, and B. Jiang, “A normal behavior model based on power curve and stacked regressions for condition monitoring of wind turbines,” IEEE Trans. Instrum. Meas. , vol. 71, pp. 1–13, 2022

    2022

  24. [34]

    Accelerated evaluation of quasi-static interaction integrals via cubic spline interpolation in the framework of the peec method,

    D. Romano, I. Kovacevic-Badstuebner, G. Antonini, and U. Grossner, “Accelerated evaluation of quasi-static interaction integrals via cubic spline interpolation in the framework of the peec method,” IEEE Trans. Electromagn. Compat., vol. 66, no. 3, pp. 829–836, 2024

    2024

  25. [35]

    Accessed: Oct

    Wtg anomaly data identification and cleaning. Accessed: Oct. 2024. [Online]. Available: https://www.datafountain.cn/competitions/451

    2024

  26. [36]

    Exploratory data analysis,

    Tukey and J. W, “Exploratory data analysis,” Reading/Addison-Wesley, 1977

    1977

  27. [37]

    The birth of mathematical morphology,

    G. Matheron and J. Serra, “The birth of mathematical morphology,” in Proceedings of the 6th International Symposium on Mathematical Morphology, pp. 1–16, 2002

    2002

  28. [38]

    Mammadov

    E. Mammadov. Wind turbine fault identification using machine learning techniques applied to scada data. Accessed: Oct. 2024. [Online]. Available: https://github.com/GCShawn/WT fault identification.git

    2024

  29. [39]

    Accessed: Dec

    Wind farm - induction generator. Accessed: Dec. 2024. [Online]. Available: https://www.mathworks.com/help/sps/ug/wind-farm-ig.html

    2024

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.