Digital Twin of Microgrid for Predictive Power Control to Buildings

Round 1

Reviewer 1 Report

The topic is very interesting, and the subject is worthy of investigation with high novelty. However, two notable issues need attention. Firstly, 

the paper's structure requires improvement, necessitating a clearer presentation of methodology, results, and discussion. Secondly, all the plots need to be improved regarding font sizes, units, axes, and image quality. Addressing these issues will significantly enhance the overall quality and readability of the manuscript.

The authors should also elaborate on the following issues before publication.

1. In line 92, the word 'thermal comfort' is used in the text. The authors need to be cautious when applying this as the definition/standard for thermal comfort may vary for different regions.
2. In the introduction section, the authors referenced three articles showcasing various models for predicting the cooling/heating loads. However, those are very lengthy and require concise summarization. The literature review should briefly summarize the previous studies' advantages and disadvantages to support the current study's objectives. While the authors have stated the purposes of their study, it is essential to provide further clarification on the research gaps.
3. The authors should elaborate on the methodology in a separate section before showing the results in Figure 1. Also, the authors should explain what AI model or algorithm is used to generate the results.
4. In Figure 1, the authors used a set of diagrams to demonstrate the overview of the Digital Twin/predictive system. However, clarity could be enhanced if the authors could explain how these results were obtained and how the digital twin validation results were compared. In addition, the subplots in Figure 1 need units for the axes, and the axes range needs to be unified. Clear definitions for each plot are also necessary.
5. In Figure 7, the font sizes in the subplots are very small. Also, the authors need to elaborate more when comparing the forecast data to observed data. Although the trends show close matches, some peaks show significant differences.
6. In Section 3.1.1, the authors need to justify the criteria applied when selecting the best model.
7. For Figure 8, missing units and legends on the Y-axis, along with the yellow line possibly obscuring the blue line, require attention. Image quality in the subplots should be improved.
8. Section 3.2 has no discussions or explanations for the results presented in Figure 9.
9. In Figure 10 and Figure 11, several data points do not match each other and show opposite trends. The authors need to explain the reasons and justify the robustness of the prediction.
10. In Section 4, the authors should include a brief summary of constraints on the current AI model used in this research, specifically regarding its accuracy.

Author Response

Dear Reviewer,

Thank you for your feedback and comment. The authors have taken an effort to address your concerns and looking forward to your favorable response. Our reply is as attached in the pdf "Reply to Reviewer 1"

Sincerely Yours,

Chew Beng 

Author Response File: Author Response.pdf

Reviewer 2 Report

1-While the LSTM model is deemed sufficiently accurate, the paper does not thoroughly explore its limitations. Acknowledging and addressing potential drawbacks or scenarios where the model may falter would strengthen the paper and provide a more comprehensive understanding of its applicability.

2- All the figures should have clear labels and graphs. The quality of the figures need improvement

3-The paper lacks a detailed analysis of the losses. A more in-depth discussion on the causes and potential mitigation strategies for energy losses would contribute to the paper's completeness.

4-The paper lacks a comparative analysis with existing methods or technologies in the field of predictive power control. Including a comparison with traditional approaches or alternative AI-based methods would provide context for the significance of the proposed methodology.

5-The study may have focused on specific datasets, and the generalization of results to different types of buildings or diverse geographical locations might be limited. The model's performance could vary when applied to datasets with characteristics not represented in the training data.

6-The paper mentions the use of a one-layer LSTM for prediction analysis. However, the sensitivity of the model to hyperparameters and the potential need for more complex architectures are not thoroughly discussed. Sensitivity analyses or comparisons with alternative models would provide a clearer understanding of the chosen approach's robustness.

7-The paper indicates that the LSTM model performs well regardless of the training data date (pre-COVID or during COVID). However, the study might not have explored potential temporal dynamics or shifts in power demand patterns that could affect the model's predictive accuracy over longer time spans.

8-The conclusion mentions optimization improvements based on the "builder operator’s preference" without explicitly defining or detailing these preferences. The lack of clarity on what factors constitute these preferences limits the reproducibility and broader applicability of the study.

9-The paper does not delve into the economic implications of the proposed system. The cost-effectiveness and feasibility of implementing the AI-Optimization-microgrid digital twin approach in real-world scenarios may be crucial considerations for its practical adoption.

Minor revision required.

Author Response

Dear Reviewer,

Thank you for your feedback and time taken to review our paper. The authors have taken an effort to address your concerns and looking forward to your favorable response. Our reply to your comment is as attached in the pdf file "Reply to Reviewer 2".

Sincerely Yours,

Chew Beng 

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper presents a predictive power control approach for improving the sustainability of buildings. The method involves using a Long Short-Term Memory (LSTM) neural network to forecast the cooling load, which is a major contributor to energy usage. The load forecasts are input to an optimization model that determines optimal setpoints for a battery energy storage system (BESS) in a microgrid supplying the building. A digital twin of the microgrid is used to evaluate the predicted BESS setpoints via real-time simulations.
Based on my review, I believe this paper has some merits but also significant weaknesses that undermine the quality and contribution:

On the positive side, the paper addresses an important area of developing predictive and optimization capabilities for improving power management in buildings. Leveraging AI methods like LSTM for forecasting along with optimization and microgrid simulation is a promising approach. There is a clear motivation around sustainability and tangible context based on the Singapore campus.

However, there are multiple deficiencies that should be addressed to strengthen the study:

Here are my criticisms of the paper:

1. The introduction lacks a clear statement of the research problem, objectives, and contributions (lines 40-81). What gap is this research addressing? How will the predictive power control approach enhance sustainability?

2. Section 2.1 provides details on the cooling system but lacks quantitative data characterizing the system such as cooling loads, COP, equipment capacities, etc. (lines 140-178). This data is needed to evaluate the experimental methods.

3. The explanation of the LSTM method in Section 2.2 is too brief (lines 183-203). More details are needed on the LSTM architecture, hyperparameters, training approach etc. to assess the rigor of the modeling.

4. The features used in the LSTM models LSTM-1 and LSTM-2 should be justified, rather than just stated (lines 213-218, Table 1). Why were these particular variables chosen as inputs?

5. The optimization formulation in Section 2.4 is unclear (lines 263-269). The objective functions and constraints should be explicitly defined with all terms included.

6. There is no explanation for how the microgrid digital twin model was validated against the physical system (lines 281-328). Quantitative results comparing the digital twin and physical asset responses are needed.  

7. In Section 3.1, the basis for comparing different LSTM architectures and claiming one layer is sufficient is not adequately supported (lines 399-401). Quantitative performance metrics across architectures should be analyzed.

8. The assumed costs for electricity, PV, and battery in the optimization study are not justified (lines 415-420). Sensitivity analysis should be included to determine how the optimal solutions change across different costs.

9. The simulation results in Section 3.3 lack quantitative indicators of model accuracy compared to experimental data (lines 482-547). Metrics like RMSE, percent errors etc. are needed.

10. There is no critical discussion of the limitations, assumptions, and potential enhancements for the overall predictive building power control framework presented (lines 552-566).

In summary, while the topic area and aims are worthwhile, the paper does not yet provide sufficient technical foundations, supporting analyses, or details to demonstrate the value of the predictive building control framework. Substantial improvements in modeling rigor, validation, and critical discussion are needed to give confidence in the quality and contributions of this research. The gaps identified should guide enhancements to the study.

The paper presents a predictive power control approach for improving the sustainability of buildings. The method involves using a Long Short-Term Memory (LSTM) neural network to forecast the cooling load, which is a major contributor to energy usage. The load forecasts are input to an optimization model that determines optimal setpoints for a battery energy storage system (BESS) in a microgrid supplying the building. A digital twin of the microgrid is used to evaluate the predicted BESS setpoints via real-time simulations.
Based on my review, I believe this paper has some merits but also significant weaknesses that undermine the quality and contribution:

On the positive side, the paper addresses an important area of developing predictive and optimization capabilities for improving power management in buildings. Leveraging AI methods like LSTM for forecasting along with optimization and microgrid simulation is a promising approach. There is a clear motivation around sustainability and tangible context based on the Singapore campus.

However, there are multiple deficiencies that should be addressed to strengthen the study:

Here are my criticisms of the paper:

1. The introduction lacks a clear statement of the research problem, objectives, and contributions (lines 40-81). What gap is this research addressing? How will the predictive power control approach enhance sustainability?

2. Section 2.1 provides details on the cooling system but lacks quantitative data characterizing the system such as cooling loads, COP, equipment capacities, etc. (lines 140-178). This data is needed to evaluate the experimental methods.

3. The explanation of the LSTM method in Section 2.2 is too brief (lines 183-203). More details are needed on the LSTM architecture, hyperparameters, training approach etc. to assess the rigor of the modeling.

4. The features used in the LSTM models LSTM-1 and LSTM-2 should be justified, rather than just stated (lines 213-218, Table 1). Why were these particular variables chosen as inputs?

5. The optimization formulation in Section 2.4 is unclear (lines 263-269). The objective functions and constraints should be explicitly defined with all terms included.

6. There is no explanation for how the microgrid digital twin model was validated against the physical system (lines 281-328). Quantitative results comparing the digital twin and physical asset responses are needed.  

7. In Section 3.1, the basis for comparing different LSTM architectures and claiming one layer is sufficient is not adequately supported (lines 399-401). Quantitative performance metrics across architectures should be analyzed.

8. The assumed costs for electricity, PV, and battery in the optimization study are not justified (lines 415-420). Sensitivity analysis should be included to determine how the optimal solutions change across different costs.

9. The simulation results in Section 3.3 lack quantitative indicators of model accuracy compared to experimental data (lines 482-547). Metrics like RMSE, percent errors etc. are needed.

10. There is no critical discussion of the limitations, assumptions, and potential enhancements for the overall predictive building power control framework presented (lines 552-566).

In summary, while the topic area and aims are worthwhile, the paper does not yet provide sufficient technical foundations, supporting analyses, or details to demonstrate the value of the predictive building control framework. Substantial improvements in modeling rigor, validation, and critical discussion are needed to give confidence in the quality and contributions of this research. The gaps identified should guide enhancements to the study.

Author Response

Dear Reviewer,
Thank you for your feedback and time taken to review our paper. The authors have taken an effort to address your concerns and looking forward to your favorable response. Our reply to your comment is as attached in the pdf file "Reply to Reviewer 3".
 
Sincerely Yours,
Chew Beng 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed all the comments. The revisions made to the manuscript have substantially improved its quality, and it is now suitable for publication.

Reviewer 3 Report

the paper can be accepted in present form

the paper can be accepted in present form