Fault Prognostics for Photovoltaic Inverter Based on Fast Clustering Algorithm and Gaussian Mixture Model
Round 1
Reviewer 1 Report
The manuscript deserves the lable "cryptic" as it discussin is mainly on not clearly defined objects. There are "features" aparently derived from moni monitoring data. We learn " t-SNE method is used for extracting two
features from 25 monitoring parameters of a PV inverter". On this method we learn: The fast clustering algorithm was based on the idea that centers are surrounded by neighbors with lower densities and characterized by large distances from points with higher densities. ". Whithout any attempt to explain what could be behind these enigmatic statements this type of information is useless. And from the figures weit can be taken that the features have numerical values ranging from -500 to + 500 [withot unit]
Shortly: without explaining what kind of numbers are derived byexactly which way from measured data and how they can be interpreted the manuscript is unintelligible.
In addition it has to be explained by which way the method described gains its "prognostic" capabilities.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
Dear Authors,
Find below my comments:
- Line 38, sudden failure should cause sudden shut down in PV system ?
- Last paragraph in the introduction should have a statement that describe clearly what is the research gab found in the literature that your paper work focused on. Also in the same paragraph need to mention what kind of verification to the proposed model was made, i.e., with other work or with experimental results.
- Line 252, need to clarify which equation you combine eq 27 with
- In section 4.1 Need to mention clearly what kind of experiment was curried and some details about the inverter type and capacity are necessary.
- Line 263 does fig 4 for maximum power output or Normalized output power
- Figure 4, must explain how did you calculate the Normalized output power.
- Figure 7, Need to show what is the difference between these 9 inverters are there difference in the specification. If there are difference then need to put it in a table.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
In this paper, an analysis is presented of a fault detection method or as the authors call it “fault prognostics” by comparing the performance of one inverter to a group of inverteres, using a mix of dimensionality reduction, clustering and a gaussian mixture model. A set of statistical tools that have previously been applied to other technologies is used to detect faults in PV inverter operation.
From an academic point of view, the paper gives an interesting analysis/method but in its current state it is not acceptable for publication. Extensive editing of English language and style is needed to improve readability, clarity and comprehensibility of the paper.
Presentation of the paper needs to be improved, including text and figures. The description of the applied methods is vague/abstract, and is suffering from unclear sentence structure and grammatical errors. The theory of the methods is shown, but how they are practically deployed is not clear. Figures are of low quality, with labels being cut-off sometimes, font size and style are inconsistent, resolution seems low, and readability of the figures is not optimal, especially e.g. figure 5, 7, 9.
In the results, finally it becomes clear what the practical approach is to determine a fault, e.g. the “warning line” threshold. The value of this threshold is only briefly discussed. This should be part of the methods, and discussed to a greater extent as it essentially determines the outcome of the algorithm. In Figure 11, a combination of Overlap Rate and JSD is given. If I understand correctly, here the Overlap Rate (rather than JSD as before) is used to signal a fault. Since Overlap = 1 – JSD, why are both shown here, it only adds to the confusion.
The conclusion is very short, and a discussion of methods and results is completely lacking.
Some other comments:
The description of input data in 2.2 seems unnecessarily abstract/vague, and equation 3 seems a function of d and k, but is not clearly correctly shown like that. What is the meaning of the power of T in eqn 2? Why do you make a “data bundle” in eqn 3 with separate days?
The relation between input data and the output of the t-SNE is not clearly discussed. The t-SNE method is discussed only in general terms and the authors fail to clearly show what comes out of this dimensionality reduction. Of the 25 features, how many remain? Do they sufficiently describe the real life operation of the inverters?
The explanation in 3.4 is not clear at all. Futhermore, the practical implementation is not shown, e.g. how all these equations lead to detection of a fault.
Figure 6: if is given as in eqn 11/12, you would expect it to have values that are multiples of 1. This is not the case, but nowhere is shown or discussed which (cut-off or gaussian) version of
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The manuscrit shows still show unexolained topics. What is exactly shown in the distribution plots (e.g. fig4).
What is the meaing of the information that a feature has a value of e.g.-750 ?. How aere these numerical values generated from exactly what data ? Why one has to deal with a symetric distributions ? No boundary effects?
There is a reasonable describtion of a technique, but an insufficient descrition how "real world data a passed through this scheme.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Thanks to the authors for the changes made. Below some remaining/further comments.
The addition of the Figure 1 is helpful for the introduction, and to understand the general approach. A part of the introduction feels like too much of a list of papers cited. It would be beneficial for the reader to summarize the key points of this literature review in a couple of sentences. E.g. this way it would become clearer what the current state of the art is how the “baseline health”, as the authors mention a few time, is determined in existing studies, and other ways this study improves the state of the art.
The methods have been improved by addition of Figures 4 and 5, but still I feel the authors need to make a better connection between the theory discussed (by means of extensive lists of equations) and the practical application in this paper. Especially section 3.4 is a full page of equations, without any mention of PV.
I understand that the quality of figures might suffer from the PDF conversions. Still, font size and style are inconsistent, and in some figures the font is just too small.
Conclusion and discussion have not improved enough. A discussion of methods, comparison with similar studies, strong and weak points of this study, etc etc is necessary. Some discussion is added, in my opinion in the wrong place (the results section lines 454-478), but this only lists advantages, and refers to drawbacks of other studies without citing any. The addition of a few lines in the conclusion is in my opinion also not enough.
The main issue with this paper is still the English language. I strongly suggest to have the paper proofread by a native English speaker. There remain many grammatical errors in throughout the paper, and the style is often colloquial.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 3
Reviewer 1 Report
The manuscript would benefit from an inntriduction to t-Distributed
stochastic neighbour embedding or adding basic literature on this topic to the references.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
No further comments.
Author Response
Thank you!
Many thanks to all those who have taken the time to review this manuscript and provide valuable comments.