the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 21 May 2025
Evaluation of Predictive Models for Reducing Wind Turbine Power Converter Failure Downtime for a Wind Farm Operator Using SCADA Data
Abstract. To facilitate the continued growth of offshore wind farm developments, operations and maintenance (O&M) costs, which are estimated at 30 % of the lifetime costs of wind farms must be reduced. This could be achieved by moving current maintenance strategies to a predictive strategy. Predictive strategies use the turbine monitoring data to determine component remaining useful lifetimes, predict failure windows or detect drops in performance and then provide an optimised maintenance plan. To enable these strategies in practice, failure prediction models must be developed, that are useable by the wind farm operator for key components. This work identifies that power converters are responsible for significant downtime at some wind farms and prediction of their failures could offer significant improvements in turbine availability. Through an analysis of their failure mechanisms, the signals required to detect failures in the power converters are identified and the insufficiencies in the SCADA data available to operators are highlighted. Several machine learning and deep learning models are trained on the SCADA to predict the power converter failures, and a novel scoring function is applied to evaluate their performance when applied to the operational decision-making context. Results suggest that implementing an artificial neural network failure prediction model offers approximately 40 % reduction in power converter maintenance costs compared to business as usual. Further improvements to these models will require the acquisition of high frequency monitoring data specific to the power electronics in the power converters. Applying predictive maintenance strategies will generate extra wind farm revenues, reduce the number of maintenance actions taken and facilitate the work of maintenance teams.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(1477 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2025-84', Anonymous Referee #1, 09 Aug 2025
The authors propose an O&M cost-based decision metric to evaluate the predictive capabilities of ML models using SCADA data that contained information on certain 16 corrective replacements for power converters for a case study wind farm. The decision thresholds, and replacement planning times were calibrated to a single case-study wind farm. The paper needs to be significantly improved with the following points that could be addressed with more clarity.
Structure and literature review
The structure of the paper is lengthy in large part from the literature review (with discussions on the types of converter faults and their locations, their causes and converter reliability statistics) that significantly digresses from the discussions on the limitations with SCADA data and the critique on machine learning models and prior work. Section 2.4 and table 2, appear to be more important from the perspective of the scope. A sharper focus on data gaps, model challenges, and their implications on predictive accuracy could improve coherence and relevance.
Performance metric
Optimizing maintenance needs based on deployment cost models is beneficial for integrating both condition-based scheduling with time based scheduled maintenance. However, the sensitivity of a model’s accuracy based on failure rates and false positive rates is not known especially since these are expected to be difficult to measure reliably. This could better indicate the problems with data/ the model’s prediction accuracies better. This is needed to strengthen the justification of the cost metric and can highlight robustness with the predictive models.
Choice of Models
The paper presents eight model architectures varying from Logistic Regression and Decision Trees to ANN and InceptionTime networks, using either 6 or 12 SCADA features. It is unclear why the specific 8-model architectures were chosen and how the specific SCADA features were used to train the 6-feature models.Results
The authors indicate that the performance of the models is highly variable depending on the corrective replacement data set that they were trained on, however, no detail is presented on which particular data sets were used for the comparisons in Table 7 and figure 16. The paper does not distinguish the limitations from training data set versus precision accuracy. Without data on carefully designed experiments using identical corrective replacement-training sets across all models, it’s hard to conclude the limitation is purely in the data. It is unclear whether the authors have tried all data mining efforts including transfer learning approaches.
What is the actual maintenance model adopted by the wind farm operator for the corrective replacement scenarios that were studied and how do the models compare in terms of the actual model adopted by the operator. Using the proposed “Business as usual (BAU)” model where no false replacements were made and all failures were missed seems to be an extreme scenario. How do these models compare against the CNN predictions made by Xiao et al., 2021 for an operational wind farm using SCADA data? Including a direct comparative analysis would better position the work within the current state of the art.
Failure Type Identification
The presented ML models and their prediction capabilities are generalized and obscure the selectivity of models in identifying specific type of corrective replacement(tied to a particular component/failure) in the power converters which will add more value for O&Os. They also do not tie back to the literature review. Component-level prediction could enhance the practical utility of the models and enable more targeted maintenance decision-making.
Citation: https://doi.org/10.5194/wes-2025-84-RC1 -
AC2: 'Reply on RC1', Demitri Moros, 12 Aug 2025
We would like to that the reviewer for their detailed comments and valuable feedback. We will be addressing these comments with work in a future publication; however, we will be withdrawing this publication due to limited resources to work on it at this time. We would like to offer a brief reply and clarification to some of the comments below.
"Structure and literature review
The structure of the paper is lengthy in large part from the literature review (with discussions on the types of converter faults and their locations, their causes and converter reliability statistics) that significantly digresses from the discussions on the limitations with SCADA data and the critique on machine learning models and prior work. Section 2.4 and table 2, appear to be more important from the perspective of the scope. A sharper focus on data gaps, model challenges, and their implications on predictive accuracy could improve coherence and relevance."
Thank you for this comment, future work will make this more focused.
"Performance metric
Optimizing maintenance needs based on deployment cost models is beneficial for integrating both condition-based scheduling with time based scheduled maintenance. However, the sensitivity of a model’s accuracy based on failure rates and false positive rates is not known especially since these are expected to be difficult to measure reliably. This could better indicate the problems with data/ the model’s prediction accuracies better. This is needed to strengthen the justification of the cost metric and can highlight robustness with the predictive models."
The failure rates were calculated from the data at the specific wind farm, it is easy to identify this failure rate as the failures of the power converters are recorded by the wind farm. The false positive rate was calculated by testing the models on data from turbines which experienced no failures across their lifetime and identifying how many replacements would have been made using the model outputs. It is agreed that measuring false positive rate reliably is very difficult but in the absence testing the models in a live environment we deemed this to be the best that could be done. Whilst failure rates will vary between wind farms, this is a number that could be calibrated on a per wind farm basis, furthermore we don't have access to data from more than one wind farm to run any sensitivity analyses.
"Choice of Models
The paper presents eight model architectures varying from Logistic Regression and Decision Trees to ANN and InceptionTime networks, using either 6 or 12 SCADA features. It is unclear why the specific 8-model architectures were chosen and how the specific SCADA features were used to train the 6-feature models."It was unclear to us if there was a specific model architecture that would perform the best with the data available. Therefore we decided to try a variety of model architectures which all work in different ways to see which would work best. The SCADA features were selected from all that were available, based on those which were thought to provide the most relevant information about the power converters, although as highlighted in some cases this was limited. The 6 feature models were trained in exactly the same way as the 12 feature models, just with fewer input features.
"Results
The authors indicate that the performance of the models is highly variable depending on the corrective replacement data set that they were trained on, however, no detail is presented on which particular data sets were used for the comparisons in Table 7 and figure 16. The paper does not distinguish the limitations from training data set versus precision accuracy. Without data on carefully designed experiments using identical corrective replacement-training sets across all models, it’s hard to conclude the limitation is purely in the data. It is unclear whether the authors have tried all data mining efforts including transfer learning approaches.
What is the actual maintenance model adopted by the wind farm operator for the corrective replacement scenarios that were studied and how do the models compare in terms of the actual model adopted by the operator. Using the proposed “Business as usual (BAU)” model where no false replacements were made and all failures were missed seems to be an extreme scenario. How do these models compare against the CNN predictions made by Xiao et al., 2021 for an operational wind farm using SCADA data? Including a direct comparative analysis would better position the work within the current state of the art."
Table 7 and Figure 16 come from test datasets that were kept separate from all the training data and so all of the models were trained and tested on the same data. We will add more description to ensure that this is more clear.
A number of different data mining efforts have been made, transfer learning was not considered, however a balance between effort expended on this research for potentially small gains and the resources available to us has meant that we haven't been able to try an exhaustive search techniques.
The actual maintenance model adopted by the wind farm being used for this case study is to fix power converters on failure. They do not have a reliable early warning system therefore they only replace failed components.
A comparison with the models used by Xiao et al. would be an excellent addition; however, we do not have access to the data that they have used and the SCADA at our case-study wind farm does not contain all of the features that they have used to train their model, therefore a direct comparison is difficult.
"Failure Type Identification
The presented ML models and their prediction capabilities are generalized and obscure the selectivity of models in identifying specific type of corrective replacement(tied to a particular component/failure) in the power converters which will add more value for O&Os. They also do not tie back to the literature review. Component-level prediction could enhance the practical utility of the models and enable more targeted maintenance decision-making."
The authors agree, component level prediction would be a significant improvement over the SCADA data, however, the data available to the operator and owner of the wind farm does not provide component level measurements. We have tried to highlight this deficiency as a recommendation for the industry to make this information available such that better predictions can be made.Once again we would like to thank the reviewer for the time spent and efforts. The feedback given will be used to make improvements in this research in future work, at this moment in time the author's are; however, withdrawing this submission.Citation: https://doi.org/10.5194/wes-2025-84-AC2
-
AC2: 'Reply on RC1', Demitri Moros, 12 Aug 2025
-
AC1: 'Comment on wes-2025-84', Demitri Moros, 11 Aug 2025
The author's are withdrawing this submission
Citation: https://doi.org/10.5194/wes-2025-84-AC1
Interactive discussion
Status: closed
-
RC1: 'Comment on wes-2025-84', Anonymous Referee #1, 09 Aug 2025
The authors propose an O&M cost-based decision metric to evaluate the predictive capabilities of ML models using SCADA data that contained information on certain 16 corrective replacements for power converters for a case study wind farm. The decision thresholds, and replacement planning times were calibrated to a single case-study wind farm. The paper needs to be significantly improved with the following points that could be addressed with more clarity.
Structure and literature review
The structure of the paper is lengthy in large part from the literature review (with discussions on the types of converter faults and their locations, their causes and converter reliability statistics) that significantly digresses from the discussions on the limitations with SCADA data and the critique on machine learning models and prior work. Section 2.4 and table 2, appear to be more important from the perspective of the scope. A sharper focus on data gaps, model challenges, and their implications on predictive accuracy could improve coherence and relevance.
Performance metric
Optimizing maintenance needs based on deployment cost models is beneficial for integrating both condition-based scheduling with time based scheduled maintenance. However, the sensitivity of a model’s accuracy based on failure rates and false positive rates is not known especially since these are expected to be difficult to measure reliably. This could better indicate the problems with data/ the model’s prediction accuracies better. This is needed to strengthen the justification of the cost metric and can highlight robustness with the predictive models.
Choice of Models
The paper presents eight model architectures varying from Logistic Regression and Decision Trees to ANN and InceptionTime networks, using either 6 or 12 SCADA features. It is unclear why the specific 8-model architectures were chosen and how the specific SCADA features were used to train the 6-feature models.Results
The authors indicate that the performance of the models is highly variable depending on the corrective replacement data set that they were trained on, however, no detail is presented on which particular data sets were used for the comparisons in Table 7 and figure 16. The paper does not distinguish the limitations from training data set versus precision accuracy. Without data on carefully designed experiments using identical corrective replacement-training sets across all models, it’s hard to conclude the limitation is purely in the data. It is unclear whether the authors have tried all data mining efforts including transfer learning approaches.
What is the actual maintenance model adopted by the wind farm operator for the corrective replacement scenarios that were studied and how do the models compare in terms of the actual model adopted by the operator. Using the proposed “Business as usual (BAU)” model where no false replacements were made and all failures were missed seems to be an extreme scenario. How do these models compare against the CNN predictions made by Xiao et al., 2021 for an operational wind farm using SCADA data? Including a direct comparative analysis would better position the work within the current state of the art.
Failure Type Identification
The presented ML models and their prediction capabilities are generalized and obscure the selectivity of models in identifying specific type of corrective replacement(tied to a particular component/failure) in the power converters which will add more value for O&Os. They also do not tie back to the literature review. Component-level prediction could enhance the practical utility of the models and enable more targeted maintenance decision-making.
Citation: https://doi.org/10.5194/wes-2025-84-RC1 -
AC2: 'Reply on RC1', Demitri Moros, 12 Aug 2025
We would like to that the reviewer for their detailed comments and valuable feedback. We will be addressing these comments with work in a future publication; however, we will be withdrawing this publication due to limited resources to work on it at this time. We would like to offer a brief reply and clarification to some of the comments below.
"Structure and literature review
The structure of the paper is lengthy in large part from the literature review (with discussions on the types of converter faults and their locations, their causes and converter reliability statistics) that significantly digresses from the discussions on the limitations with SCADA data and the critique on machine learning models and prior work. Section 2.4 and table 2, appear to be more important from the perspective of the scope. A sharper focus on data gaps, model challenges, and their implications on predictive accuracy could improve coherence and relevance."
Thank you for this comment, future work will make this more focused.
"Performance metric
Optimizing maintenance needs based on deployment cost models is beneficial for integrating both condition-based scheduling with time based scheduled maintenance. However, the sensitivity of a model’s accuracy based on failure rates and false positive rates is not known especially since these are expected to be difficult to measure reliably. This could better indicate the problems with data/ the model’s prediction accuracies better. This is needed to strengthen the justification of the cost metric and can highlight robustness with the predictive models."
The failure rates were calculated from the data at the specific wind farm, it is easy to identify this failure rate as the failures of the power converters are recorded by the wind farm. The false positive rate was calculated by testing the models on data from turbines which experienced no failures across their lifetime and identifying how many replacements would have been made using the model outputs. It is agreed that measuring false positive rate reliably is very difficult but in the absence testing the models in a live environment we deemed this to be the best that could be done. Whilst failure rates will vary between wind farms, this is a number that could be calibrated on a per wind farm basis, furthermore we don't have access to data from more than one wind farm to run any sensitivity analyses.
"Choice of Models
The paper presents eight model architectures varying from Logistic Regression and Decision Trees to ANN and InceptionTime networks, using either 6 or 12 SCADA features. It is unclear why the specific 8-model architectures were chosen and how the specific SCADA features were used to train the 6-feature models."It was unclear to us if there was a specific model architecture that would perform the best with the data available. Therefore we decided to try a variety of model architectures which all work in different ways to see which would work best. The SCADA features were selected from all that were available, based on those which were thought to provide the most relevant information about the power converters, although as highlighted in some cases this was limited. The 6 feature models were trained in exactly the same way as the 12 feature models, just with fewer input features.
"Results
The authors indicate that the performance of the models is highly variable depending on the corrective replacement data set that they were trained on, however, no detail is presented on which particular data sets were used for the comparisons in Table 7 and figure 16. The paper does not distinguish the limitations from training data set versus precision accuracy. Without data on carefully designed experiments using identical corrective replacement-training sets across all models, it’s hard to conclude the limitation is purely in the data. It is unclear whether the authors have tried all data mining efforts including transfer learning approaches.
What is the actual maintenance model adopted by the wind farm operator for the corrective replacement scenarios that were studied and how do the models compare in terms of the actual model adopted by the operator. Using the proposed “Business as usual (BAU)” model where no false replacements were made and all failures were missed seems to be an extreme scenario. How do these models compare against the CNN predictions made by Xiao et al., 2021 for an operational wind farm using SCADA data? Including a direct comparative analysis would better position the work within the current state of the art."
Table 7 and Figure 16 come from test datasets that were kept separate from all the training data and so all of the models were trained and tested on the same data. We will add more description to ensure that this is more clear.
A number of different data mining efforts have been made, transfer learning was not considered, however a balance between effort expended on this research for potentially small gains and the resources available to us has meant that we haven't been able to try an exhaustive search techniques.
The actual maintenance model adopted by the wind farm being used for this case study is to fix power converters on failure. They do not have a reliable early warning system therefore they only replace failed components.
A comparison with the models used by Xiao et al. would be an excellent addition; however, we do not have access to the data that they have used and the SCADA at our case-study wind farm does not contain all of the features that they have used to train their model, therefore a direct comparison is difficult.
"Failure Type Identification
The presented ML models and their prediction capabilities are generalized and obscure the selectivity of models in identifying specific type of corrective replacement(tied to a particular component/failure) in the power converters which will add more value for O&Os. They also do not tie back to the literature review. Component-level prediction could enhance the practical utility of the models and enable more targeted maintenance decision-making."
The authors agree, component level prediction would be a significant improvement over the SCADA data, however, the data available to the operator and owner of the wind farm does not provide component level measurements. We have tried to highlight this deficiency as a recommendation for the industry to make this information available such that better predictions can be made.Once again we would like to thank the reviewer for the time spent and efforts. The feedback given will be used to make improvements in this research in future work, at this moment in time the author's are; however, withdrawing this submission.Citation: https://doi.org/10.5194/wes-2025-84-AC2
-
AC2: 'Reply on RC1', Demitri Moros, 12 Aug 2025
-
AC1: 'Comment on wes-2025-84', Demitri Moros, 11 Aug 2025
The author's are withdrawing this submission
Citation: https://doi.org/10.5194/wes-2025-84-AC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 218 | 67 | 23 | 308 | 15 | 33 |
- HTML: 218
- PDF: 67
- XML: 23
- Total: 308
- BibTeX: 15
- EndNote: 33
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1