Two-Layer Ensemble-Based Soft Voting Classifier for Transformer Oil Interfacial Tension Prediction

Round 1

Reviewer 1 Report

1. What is the new contribution of this paper comparing with ref. [7]? Please build up the new contribution in Introduction.
2. I think the proposed method is simple combination of eight difference classfiers. What are the new methodology or features of proposed method?
3. I cannot find the advantages of proposed method in section 2.3. Please revise section 3 to present the advantages of proposed method.
4. Conclusions are very short.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

In Figure 1. The data distribution of the (a) trianing data set and (b) testing data set

"trianing" perhaps you want  to say "training"

You must write in all text “dataset” better than “data set”

------------

You used 730 transformers 66 kV, 12.5-40 MVA  in the training of the learning machine. That kind of transformers is used in the high-voltage electrical transport system or in the high-voltage distribution system.

It is compared with 36 transformers, 13.8 kV, 0.5-1.5 MVA. That kind of transformer is used in medium voltage distribution systems.

High-voltage transformers have a different failure rate than medium-voltage transformers. Short-circuit faults and overheating are more frequent in medium-voltage transformers than in high-voltage transformers. Oil and paper have different aging curves.

Can this be a drawback in the proposed learning model?

 

 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

Paper presents proposal of classification method for prediction of the interfacial tension of the transformer oil using 5 predictors (other oil parameters). Authors used several machine learning methods and tried to optimize the number of input parameters – as a result two layer ensemble based method gave the best accuracy of the prediction (using all 5 predictors). The raised problem is quite interesting and up-to-date, especially regarding the general technical condition assessment (and its prediction) of power transformers. Nevertheless, in the reviewer opinion paper has several flaws and confusing issues that should be improved in order to make the research clear for the reader. Thus some key notes are listed below.  

1. According to the title one expects that paper is dealing with “transformer insulation assessment” since it isn’t. Of course abstract explains it in the adequate way, but title is misleading, in the reviewer opinion. Generally, paper proposes a classifier for ITF only – not considering its impact on the insulation assessment at all. Maybe authors try to consider revision of the title – to make it more adequate.
2. Authors claim that “ITF is a very important parameter that needs to be evaluated to assess the transformer oil condition”, but they do not comment or justify it in the paper. As ITF is the main focus of the paper it should be explained for the reader what information it brings about the oil condition.
3. Reviewer has real doubts about the relevance of data used for classification. Authors used data form 66kV units (mid power) to train models, and then data from 13kV (low power) units to validate them. It is quite questionable, since those are completely different transformers – not only according to their voltages, but also construction, cooling, load profile etc. Which means that they usually have different requirements regarding diagnostic parameters (also the oil properties). For example in the reviewer region for units >2.5MVA allowed ITF is >24, while for units <2.5MVA it is >20, or is not analyzed at all. The most surprising is that authors know that and highlight it in the text – but, unfortunately for the reader, not justify their choice in a rational way. Models accept any data for training and always give some results – even if data is selected wrongly accuracy may be relatively good. This issue should be discussed.   
4. According to the reviewer, the output of the paper would be much more relevant if authors supported their results by several case studies – where real measurement data would be shown and analyzed. For example, consider to show real data set (input and output) for 3 exemplary cases: unit for whom the accuracy was high, unit with moderate accuracy and with poor accuracy. Results should also be extended with confusion matrix (target class vs output class), showing number of units classified correctly and incorrectly.
5. According the state-of-art there is a strong relation between acidity and ITF (Acidity raises and ITF falls when oil oxidation develops). Authors’ results showed that its importance is rather moderate – it should be commented.
6. Reviewer feels that correlation matrix should be extended with ITF – despite it is not an input parameter it would show which parameters are most closely correlated with it. It could also support the feature importance results (fig.5).
7. Why to show potential costs of the oil tests using sources from 10 years ago? It is a little bit confusing. Especially, since such costs vary a lot, even between different countries.
8. Reference list is very modest – transformer diagnostics or machine learning are very common topics. A more detailed discussion of the state of the art is expected.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I have no further comments.

 

Author Response

Thanks for your feedback.

Reviewer 3 Report

Authors made some enhancement of the paper comparing it to its previous version. In general, authors responded to most of the reviewer’s suggestions and doubts (in cover letter only) - unfortunately, only few of them are also reflected in the revised paper. In the reviewer’s opinion several further improvements are highly recommended, as follows.

1. Authors should note that potential readers do not see the response letter (contrary to the reviewer) – which means that if they have similar questions they will not find the explanation for them in the paper. In fact, reviewer is the very first reader of authors’ paper, and it is highly probable that other readers may have similar doubts/questions. Thus, all of the explanations to the reviewer doubts (if any) should be included in the paper also – not in the response letter only (especially when they are strictly related to the text). This generally refers to the answer to previous note 3, which is not included in the revised paper.
2. Authors did not respond to the main part of the previous note 4 – once again: in the reviewer opinion authors should support their results by several case studies – where real measurement data would be shown and analyzed. For example, consider to show real data set (input – measurements results, and output – diagnosis/) for 3 exemplary cases: unit for whom the accuracy was high, unit with moderate accuracy and with poor accuracy. It would be very useful information for the readers, especially from the practical point of view (as the authors intent to propose a method for real-life transformers).
3. A general hint - According to the answer for the previous note 3, authors claim that “The difference in the limits on the acceptable value of IFT is imposed due to the importance of the transformer”. It is only partially true. The real reason (from the scientific point of view) is the electric stress (voltage level) - which in fact is usually strictly related to the unit’s importance. Higher voltage means higher requirements for the insulation system, because probability of the potential insulation failure due to the electric stress is higher also.

Author Response

Please see attached.

Author Response File: Author Response.docx

Round 3

Reviewer 3 Report

Analyzing the authors’ answers, reviewer came to the conclusion that one of his notes was imprecise – note 2 form the previous report. Sorry about that. The intention was to encourage authors to show some exemplary values of the dataset. As authors mentioned in the paper “The input features included in the dataset are water content, acidity, breakdown voltage, dissipation factor (DF) and color while the output variable is the interfacial tension (IFT)” – please show exemplary (input) values of water content (e.g. 7ppm), acidity (e.g. 0.08 mgKOH/g_ol), breakdown (72kV) etc. vs  IFT (e.g. 25 dyne/cm) for the relevant datasets with respect to 3 scenarios of IFT: good( e.g. 35 dyne/cm), moderate – close to the boundary value (e.g. 30 dyne/cm), bad (e.g. 21 dyne/cm). Hope it is clear now. All other reviewer’s comments have now been addressed in the revised manuscript.

Author Response

In my research I solved the machine learning problem as a classification problem and not as a regression problem. It is a two-class problem where the output is either good or bad, i.e. two classes. In regression, we predict the actual value but I did not do that in my research. Regression is totally a different problem where you need to train the data using actual value of IFT and not just a class.