Research on Power Flow Prediction Based on Physics-Informed Graph Attention Network

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper addresses the problem of microgrid power flow prediction, by describing the limitations of traditional statistical and time-series approaches that ignore spatial node relationships and by proposing an innovative approach based on AI techniques.

 The authors propose a Physics-Informed Graph Attention Network (PI-GAT), which combines an attention mechanism with physical state equations integrated into the loss function.

The integration of spatial attention into graph convolution is the key that allows to improve the modeling of node interactions and dependencies within microgrids, leading to more accurate power flow predictions.

Comments

The paper is well-written and well-structured, and it presents an interesting and timely approach to a problem of growing importance in distributed energy systems. 

Introduction

• Line 63: “attention network (PI-GAT)is proposed” → there should be a space after the bracket: “(PI-GAT) is proposed”.
• Line 75: “We constructa comprehensive…” → should be corrected: “We construct a comprehensive…”.

Section 2: Basic Theory

• Subsection 2.1 (Lines 116–118): the meaning of an edge should be clarified, like:
  “An edge is usually defined as a pair of nodes from the node set VVV, i.e., ei=(vj,vk)e_i = (v_j, v_k)ei​=(vj​,vk​) with vj,vk∈Vv_j, v_k \in Vvj​,vk​∈V.”
• Figure 1: the origin of the image should be indicated (it is original or adapted from another source?). In general, it could be useful to indicate when a figure is taken from another paper.
• Line 170: “The specific structure is shown in Figure 3.” → specify Figure 3a.
• Line 179: “The principle of single-layer attention is shown in Figure 3.” → specify Figure 3b.
• Line 258: “N(i)represents…” → add missing space: “N(i) represents…”.

Section 5: Experiments

• Were the  hyperparameters set manually or tuned systematically?. The authors should clarify the procedure.
• Line 450: in Figure 9, the reference is not hyperlinked (appears in black instead of blue). This should be fixed.
• Line 479: the sentence about node representations needs better phrasing, like:
  “Graph convolution layers update node representations by convolving each node’s features with those of its neighbors, effectively capturing the graph's structural information.”
• Subsection 5.3.3 (Line 520): the reference to “Figure 12” is incorrect. It should be “Figure 10”.

Section 5.7: Conclusion

• Line 573: the sentence “By conducting comparative experiments and analyzing the results, the following conclusions can be drawn:” should be followed by bullet points. Otherwise the colon should be replaced with a period.

Acknowledgments

• This section should either be removed or properly completed.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

• Quantify the improvement in using PI-GAT in abstract
• L22: Cite "energy Internet"
• L59: GCN are not always bad but one can try GCN with weighted adjacency which does the same thing as GAN
• L62: You need to also explain how you want to incorporate physics loss (PI) before saying PI-GAT
• L75:  "contructa" -> "construct a"
• Figure 2: A nice fully connected AE diagram can be placed instead of the current one
• Figure 7(a): please correct the caption

> The main question addressed by the research: Can a physics-informed graph attention network (PI-GAT) improve the accuracy and robustness of microgrid power flow prediction compared to traditional methods and other deep learning architectures

> The topic is relevant to the field of smart grids, energy management, and AI for power systems, since accurate power flow prediction is essential. The proposed PI-GAT addresses the gap (from graph and attention NN) by introducing attention and embedding physical equations in the loss function. And a very good integration of the power loss functions in the training the model.

> Other work that is prevailing in the field used GNN or transformer models but in this paper, systematic comparisons against baselines (DNN, CNN,GCN, etc.) showing better performance of the proposed model

> I found that the existing methodology is robust but I can see the best possibilities combined (Auto encoder + attention + physics loss function), one can compare where most of the advantage comes from....

But this would go in supplementary/appendix. They already mention that this is a proof of concept for 14 microgrid architecture, but the methodology would of course adapt with scaling.

> The plots and tables show clear improvement in accuracy and reduced loss, which justifies the abstract. But again this is done on a small scale setup

> All the references to GNN, power systems, attention NN, PINN are cited properly.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper focuses on the development of a Physics-Informed Graph Attention Network (PI-GAT) aimed at improving power flow predictions in AC microgrids. AC Microgrids play a relevant role in improving energy dispatch and grid operations. It seems that prediction techniques often fail to consider the spatial relationships between the nodes within these systems. The PI-GAT addresses this gap by employing an attention mechanism that evaluates the significance of connections between nodes, thereby enhancing the accuracy of power flow predictions.

The work on power flow prediction using the PI-GAT model presented in this research represents an interesting step forward in the field of energy management in AC microgrids. By combining physics-informed approaches with advanced neural network architectures, the study enhances predictive capabilities and operational efficiency in microgrid systems, paving the way for future innovations in smart grid technology.

Also,

The abstract does not include a critical comparative analysis of relative advantages.

Authors state in the abstract “Finally, simulation verification was conducted, comparing the PI-GAT method with traditional approaches.” What about comparing the proposal to new approaches. It is expected that any newcomer solution outperforms "old" traditional approach. What is the case if they do not? Maybe the reason behind the lack of comparing to new proposals is that other new proposals were lightly reviewed.

The manuscript is too wordy. At least sections 2 and should be more simplified, more concise. An example of this is that the first paragraph in section 3.1 8 (Due to the access of……a typical 14-node microgrid structure diagram) is not necessary to understand the idea. If the whole manuscript is rewritten to get more concise paragraphs, the better for readers.

Figure 3 fits better before section 3

Review the whole manuscript for typos and fix them.

Is the microgrid in Figure 4 an original proposal from the Author or does it need a reference? What should be understood as “typical 14-node microgrid structure diagram.” Is this a IEEE proposal?

It must be emphasized that the work is about AC microgrids (DC are excluded in this work).

Based on which criteria the topology in Fif. 5 was broken in the topologies shown in Fig. 5?

What should be understood as “industrial scenarios”?  Please give some examples

Can the PI-GAT be applied to other areas of energy management beyond microgrids?

Comments on the Quality of English Language

The manuscript need minor edition. There are various typos.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

In this manuscript, the authors propose a Physical-Informed Graph Attention Network (PI-GAT) to predict the power flow in microgrids, in particular, they studied microgrids consisted on 14-node with different network's topology.

The manuscript is well written, and can ever serve as a didactic manuscript given the level of detail of the explanations. However, I found surprising that there are only references to the literature up to page 9 (of 20), something surprising even in Section 5.3.1. Model Comparison, where the comparison with OTHER models is carried out. In my opinion, this fact reduces the quality of the manuscript, and makes me suspect if many relevant literature is missing.

It is also not clear to me what is the source of the training data of this model, or even if the dataset is syntetic. Please, clarify this point in the text.

Therefore, I ask the authors to add more relevant references to their manuscript and answer the question about the source of the data used, so I can reconsider my recommendation.

Comments on the Quality of English Language

I have found several typos:

Line 75: Typo: replace "constructa" by "construct a"
Line 128: Typo: "metrix" by "matrix"
Line 128: Typo: "adjaceny" by "adjacency"
Sentence in lines 349-351: Probable typo in in low-dimensional data x_x, is it F' or F?
Table 3: Is really the number of layers of Type AE_3 1 instead of 3?
Table 6: Typo: "Propotion"

Therefore, I ask the authors to check the english spelling of the article to correct typos

Author Response

Please see the attachment.

Author Response File: Author Response.pdf