A Neural Network Approach to Physical Information Embedding for Optimal Power Flow
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this paper, the authors proposed an architecture using physics informed neural network to optimize power flow in a power system. Overall, the paper provides clear description on the neural network method as well as sufficient results to support. This manuscript should be eligible for publication after minor revisions.
1. There seems to be a typo at line 140, where constraints (2)-(5) are equality constraints. Also, S seems to represent all parameters in the set in (17) rather than just current. Some clarifications are needed here.
2. The authors described the loss function in general physical information architecture in section 2.2 and the loss function they used in training in section 3.3. While the physics informed part is clear, there seems to be terms that are not reflected in section 3.3 and some clarifications may help understand better.
3. line 277 seems to be missing a reference.
4. Table 4 has a title of training time but it is describing worst-case guarantee. This is confusing and some clarification is needed on what is discussed here.
Comments on the Quality of English LanguageEnglish is fluent and clear. Minor typos need to be corrected.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses in the attachment and the corresponding revisions in the re-submitted files. Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper proposes a new framework for AC optimal power flow analysis based on a physically-informed convolutional neural network approach. This study also develops a method for extracting and minimising the worst-case generation constraints violations for the PICNN. Experimental results demonstrate that the method to be highly accurate and robust for the problem of AC optimal power flow analysis problem.
The paper presents sufficient theoretical material, and the methods are adequately described. A comprehensive study has been carried out and the experimental results are clearly presented. The references are appropriate.
I recommend the acceptance of this paper after some checking of the text for formatting inaccuracies.
Author Response
Dear reviewer,
Thank you very much for taking time out of your busy schedule to review this manuscript. We have carefully checked and revised your questions about the text format, and revised the manuscript as required in the follow-up process. Thank you again for your review and feedback on our paper, and thank you for affirming our research content.
Sincerely,
Chenyuchuan Liu
Reviewer 3 Report
Comments and Suggestions for AuthorsThere are some problems, which must be solved before it is considered for publication. If the following problems are well-addressed, this reviewer believes that the contribution of this paper are important.
1. The schematic diagram of the network framework in the paper is too blurry to clearly see the relevant parameters and design model. Please provide the corresponding vector graphics to resolve this issue.
2. The author repeatedly states that the proposed model can reduce dependency on datasets, but this is not substantiated in the paper. Could the author provide some data or formulas to demonstrate the model's contribution to reducing dataset dependency?
3. The paper mentions that the initial weight parameter matrix has a significant impact on the final results. How did the author ultimately select the optimal parameter matrix? Could some specific values of the weight matrices be provided to enhance persuasiveness? Additionally, in the final experimental comparison section, were the LSTM, PINN, and PICNN models all using optimized initial parameters, including weight matrices and hyperparameters?
4. In the design of the network structure described in the paper, what are the specific parameters of the network, such as the size of the convolution kernels, stride size, and whether padding is used? How is the PRT component of the loss function determined, and how does PRT function within the network's feedback mechanism?
5. The paper mentions the use of max pooling to reduce data dimensionality. Does this lead to data loss issues? Would using average pooling mitigate this problem?
6. During the training process, what are the formats and dimensions of the input and final output data of the network? When constructing the dataset, what are the original data and their corresponding labels? The paper previously mentioned a dataset split ratio of 7:3, so why is a 5:2:3 ratio used to divide the validation, training, and testing sets in the "Results" section?
Comments on the Quality of English LanguageIt is recommended to have the English language polished by a professional organization
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses in the attachment and the corresponding revisions in the re-submitted files. Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
Comments and Suggestions for AuthorsA paper is proposed on improving the efficiency of power flow using neural networks. Currently, neural networks are very popular and their application is found in the energy sector, among other things. Such solutions allow you to speed up the calculation process and find a more optimal solution in the shortest possible time. Nevertheless, network training does not always go fast, correct algorithms are needed here.
The authors managed to achieve the desired results in comparison with existing solutions. They considered and proposed the optimal distribution of the power flow, which was confirmed by theoretical calculations. The following is the architecture of the neural network, where the necessary network learning algorithms are laid down.
Each such study is unique, and therefore carries novelty and a deposit for future research.
Nevertheless, a non-neural processor was used for the experiment, which probably did not give the best results, nevertheless they were obtained and presented in an accessible way.
The results and conclusions fully correspond to the set goal and objectives, where there was a search for optimal power distribution, which the authors achieved.
The links to which the authors refer fully satisfy the subject of the article.
There is a small note on the design – the quality of the drawings, Figure 1, Figure 2 are very small and difficult to read, you need to increase the font inside the drawings. In general, the work can be accepted for publication.
Author Response
Dear reviewer,
Thank you very much for taking time out of your busy schedule to review this manuscript. For your questions about the quality of the drawing, we have replaced it with a clearer vector drawing. Thank you again for your review and feedback on our paper, and thank you for affirming our research content.
Sincerely,
Chenyuchuan Liu
Reviewer 5 Report
Comments and Suggestions for AuthorsThe paper proposes a novel Physics-Informed Convolutional Neural Network (PICNN) approach to solve the AC optimal power flow problem.
1. Embedding physical constraints into neural networks through the loss function has been applied in other fields. Although this paper introduces physical constraints, it does not fully leverage the unique structure and characteristics of power systems.
2. The proposed PICNN architecture is fixed. How can it be proven suitable for power systems of all scales?
3. The description of the PICNN architecture in Section 3 lacks detail. It is recommended to provide more specifics, such as the parameters for each layer.
4. As mentioned in line 165 about assigning physical meaning, where is this specifically demonstrated? For example, explain the parameters in equation (26) from the physical perspective.
5. What is the unit of training time—minutes or seconds? What is the actual computation time, and can it meet the requirements for real-time application?
6. Please verify the accuracy of line 313, as MW is a unit of power.
Comments on the Quality of English LanguageThe English language in the manuscript is generally clear and comprehensible.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses in the attachment and the corresponding revisions in the re-submitted files. Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
Comments and Suggestions for AuthorsThe author has completed all the revisions and recommends acceptance for publication.
Reviewer 5 Report
Comments and Suggestions for AuthorsAlthough some of the author's responses mainly indicate that further work will address these issues, the explanations are generally acceptable.
