Review Reports - A Real-Time Photovoltaic Power Estimation Framework Based on Multi-Scale Spatio-Temporal Graph Fusion

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a spatio-temporal graph learning framework (WGL) for ultra-short-term photovoltaic (PV) power forecasting. The model integrates learnable wavelet shrinkage, TCN-LSTM fusion, and graph attention mechanisms. While the topic is timely and the model demonstrates competitive performance, the manuscript suffers from limited novelty, insufficient experimental validation, and lack of deployment details. Major revisions are needed to meet the publication standards.

There are a large number of irregularities in the writing format, e.g., Figure.1, Figure.2, etc., commonly Figure 1 or Fig. 1. In addition, the initial letters in the titles and captions are not capitalized and are not standardized.
Figures 9, 10 and 11 are not sufficiently clear and do not distinguish the colours significantly.
The dataset is limited to three PV stations in Greece, which is insufficient to demonstrate generalizability. Inclusion of multi-region or multi-climate datasets is strongly recommended.
The ablation study lacks fine-grained analysis of spatial modules (GSA and FSTA). It is unclear how much each contributes to the final performance.
There is no uncertainty quantification (e.g., prediction intervals, Bayesian modeling), which is critical for real-world power system applications.
The manuscript contains numerous grammatical errors and unprofessional expression. A thorough proofreading is recommended.
The manuscript claims to support "real-time" forecasting, yet no computational complexity, inference time, or memory usage is reported. These metrics are essential for evaluating deployment feasibility.

Author Response

Reviewer 1:This paper proposes a spatio-temporal graph learning framework (WGL) for ultra-short-term photovoltaic (PV) power forecasting. The model integrates learnable wavelet shrinkage, TCN-LSTM fusion, and graph attention mechanisms. While the topic is timely and the model demonstrates competitive performance, the manuscript suffers from limited novelty, insufficient experimental validation, and lack of deployment details. Major revisions are needed to meet the publication standards.

There are a large number of irregularities in the writing format, e.g., Figure.1, Figure.2, etc., commonly Figure 1 or Fig. 1. In addition, the initial letters in the titles and captions are not capitalized and are not standardized.
Figures 9, 10 and 11 are not sufficiently clear and do not distinguish the colours significantly.
The dataset is limited to three PV stations in Greece, which is insufficient to demonstrate generalizability. Inclusion of multi-region or multi-climate datasets is strongly recommended.
The ablation study lacks fine-grained analysis of spatial modules (GSA and FSTA). It is unclear how much each contributes to the final performance.
There is no uncertainty quantification (e.g., prediction intervals, Bayesian modeling), which is critical for real-world power system applications.
The manuscript contains numerous grammatical errors and unprofessional expression. A thorough proofreading is recommended.
The manuscript claims to support "real-time" forecasting, yet no computational complexity, inference time, or memory usage is reported. These metrics are essential for evaluating deployment feasibility.

Response(1)

Thank you very much for your helpful comment. We have carefully checked and standardized the writing format throughout the manuscript. Specifically, all figure references have been corrected to the appropriate form (e.g., “Figure 1” or “Fig. 1” depending on journal style), and the capitalization of titles and captions has been revised for consistency and compliance with the required formatting guidelines.

Response(2)

Thank you for your valuable comment. We agree that clear visual distinction is important. In our experiments, multiple models are displayed together in the same figures, and the color contrast between them has already been carefully designed to ensure differentiation. However, to further improve readability, we have enhanced the resolution and clarity of Figures 9, 10, and 11. Additionally, we have uploaded high-resolution versions of each figure to ensure that the color distinctions and details are clearly visible in the final manuscript.

Response(3)

Thank you for your constructive comment. We agree that demonstrating generalizability across different regions and climate conditions is important. In addition to the three PV stations in Greece originally presented in the manuscript, we have further conducted comprehensive generalization experiments using datasets from 18 wind farms located in multiple regions with diverse climate characteristics. The results consistently show that our proposed WGL model maintains superior forecasting performance compared with baseline models under these additional datasets. These supplementary experiments validate the robustness and generalizability of WGL across different geographical and meteorological environments. The corresponding experimental results and analysis have been added to the revised manuscript.

Response(4)

Thank you for your valuable comment. To address the reviewer’s concern regarding the fine-grained contribution of the spatial modules (GSA and FSTA), we have added a step-by-step ablation study in Table 4 of Section 3.2 in the revised manuscript. Instead of removing modules individually, we follow a cumulative ablation strategy, where modules are introduced progressively to the baseline model (Baseline → +LWS → +T-MSFF → +GSA → +FSTA → WGL). The results clearly show that both GSA and FSTA contribute positively to the final forecasting performance. Specifically, adding GSA improves RMSE/MAE/MAPE from 0.62/0.37/1.05 to 0.60/0.36/1.01, and adding FSTA further enhances performance to 0.57/0.33/0.91 in the complete WGL model. These findings demonstrate that each spatial attention module provides meaningful performance gains, while their combination yields the best accuracy.

Response(5)

Thank you for pointing out the importance of uncertainty quantification in practical power system applications. We fully agree that incorporating prediction uncertainty (e.g., prediction intervals or Bayesian modeling approaches) is essential for real-world deployment scenarios. In this work, our primary focus is on improving deterministic forecasting accuracy. However, uncertainty quantification is indeed an important direction, and we plan to extend the proposed WGL framework to support probabilistic forecasting and interval estimation in our future research. We have added a corresponding statement in the conclusion section of the revised manuscript to clarify this point.

Response(6)

Response(7)

Thank you for highlighting the need to evaluate the deployment feasibility of the proposed model. In response, we have added quantitative analyses of computational complexity and inference efficiency in the revised manuscript. Specifically, Table 6 now reports the number of parameters, MFLOPs, and inference latency for all baseline models and the proposed WGL model. As shown, WGL achieves competitive computational efficiency, requiring fewer parameters than Transformer-based baselines and demonstrating lower latency (6.15 ms), which is the lowest among all models compared in the study. These results indicate that WGL maintains strong forecasting accuracy while supporting real-time inference with efficient memory usage and computational cost. The corresponding experimental results and discussion have been added to Section 3.2 of the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The generalization capability of the proposed WGL framework requires further substantiation, as its evaluation is confined to a limited regional dataset. Validation across diverse geographical and climatic regions is necessary to robustly demonstrate the model's universality and performance stability beyond the presented case study.
The individual contribution and potential redundancy of the integrated complex modules remain unclear. A more rigorous ablation study is recommended to precisely disentangle the necessity of each component and to assess the performance-complexity trade-off relative to refined baseline architectures.
Claims of "real-time" performance lack supporting empirical evidence on inference latency and computational efficiency. Quantitative benchmarks, such as inference speed and resource consumption, are essential to validate practical deployment feasibility and to inform potential optimization pathways for large-scale applications.
The framework's exclusive focus on point forecasts limits its utility for operational risk assessment. Incorporating probabilistic forecasting techniques to quantify predictive uncertainty would significantly enhance its practical value for grid management decisions under variable conditions.
The figure formats in this article are not uniform, and some images are not rigorous enough.

Author Response

Reviewer 2:

The generalization capability of the proposed WGL framework requires further substantiation, as its evaluation is confined to a limited regional dataset. Validation across diverse geographical and climatic regions is necessary to robustly demonstrate the model's universality and performance stability beyond the presented case study.
The individual contribution and potential redundancy of the integrated complex modules remain unclear. A more rigorous ablation study is recommended to precisely disentangle the necessity of each component and to assess the performance-complexity trade-off relative to refined baseline architectures.
Claims of "real-time" performance lack supporting empirical evidence on inference latency and computational efficiency. Quantitative benchmarks, such as inference speed and resource consumption, are essential to validate practical deployment feasibility and to inform potential optimization pathways for large-scale applications.
The framework's exclusive focus on point forecasts limits its utility for operational risk assessment. Incorporating probabilistic forecasting techniques to quantify predictive uncertainty would significantly enhance its practical value for grid management decisions under variable conditions.
The figure formats in this article are not uniform, and some images are not rigorous enough.

Response(1)

We sincerely thank the reviewer for the insightful comment on the generalization capability of the proposed WGL framework. In response, we conducted additional validation experiments using newly collected data from 18 wind farms situated in diverse geographical regions and spanning multiple seasons and meteorological conditions.The comparative results, now presented in Section 3.2 and Table 7, demonstrate that the WGL framework consistently outperforms baseline models and maintains stable performance across all heterogeneous datasets. This indicates that the method is not restricted to the original regional dataset and exhibits strong robustness and generalization ability.We have also added a detailed discussion in the manuscript to further clarify the model’s universality and the practical implications of its cross-regional performance.

Response(2)

We appreciate the reviewer’s insightful suggestion. We would like to clarify that the individual contributions of each module have already been analyzed in detail in Section 3.2 of the manuscript. In this section, we conducted a comprehensive ablation study by systematically removing or replacing each component of the proposed framework, and compared the results with corresponding refined baseline architectures. The experimental results clearly demonstrate that each module plays a distinct and indispensable role in improving model performance.

Response(3)

Thank you for the reviewer’s valuable comments. To provide empirical support for our ‘real-time’ claim, we added and reported quantitative benchmarks—end-to-end inference latency, computational cost (MFLOPs), and parameter count—in Section 3.2/Table 6, covering the same models and settings as the accuracy experiments. The results show that WGL has 325K parameters, 5.12 MFLOPs, and an inference latency of 6.15, outperforming strong baselines under the same environment (e.g., Transformer: 415.75K, 7.275 MFLOPs, 9.56; CNN-LSTM: 372.77K, 6.61, 14.19). Although CNN has the smallest parameter count and MFLOPs (8.129K, 0.086), it exhibits the highest latency (15.91), indicating limited parallel efficiency; Transformer-based models show better parallelism with moderate latency, whereas WGL achieves the lowest latency while maintaining moderate model size and computational cost, demonstrating a superior accuracy–efficiency trade-off.

Response(4)

We appreciate the reviewer’s suggestion on probabilistic forecasting. We fully recognize the importance of uncertainty quantification for operational risk assessment; however, the objective and application scope of this work are targeted at the deterministic (point) forecasting stage within short-term rolling scheduling/control pipelines. In many real systems, risk mitigation is primarily handled downstream by security-constrained unit commitment, reserve allocation, and alarm policies, while the optimization layer typically requires a single-valued input to ensure determinism and meet strict latency constraints. Accordingly, this paper focuses on improving the accuracy and stability of point forecasts under fixed compute and time budgets—an improvement that directly translates into lower reserve requirements and reduced operational risk. As demonstrated across multiple sites, horizons, and baselines, WGL achieves significant gains on RMSE/MAE/MAPE. We agree that probabilistic extensions are valuable and orthogonal/compatible with our framework, but they introduce additional modeling assumptions, evaluation dimensions, and engineering overhead that lie beyond the present scope. For practitioners who need uncertainty estimates, we clarify that post-hoc methods—e.g., empirical residual distributions/bootstrap, quantile mapping, or conformal calibration—can be applied without modifying WGL to derive intervals or quantiles. In sum, we believe a point-forecast focus is appropriate for the stated goals and constraints, and we plan to explore and report probabilistic extensions in future work and releases to further support risk-aware grid decision making.

Response(5)

Thank you for your comment. We have carefully revised the figures to ensure uniform formatting across the manuscript, including consistent font sizes, line styles, and labeling. In addition, we have improved the clarity and rigor of several images to enhance their scientific accuracy and visual quality.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper tackles a pressing challenge in renewable energy integration: robust, accurate, and real-time short-term PV power forecasting at multiple, spatially distributed sites. The proposed WGL framework is innovative, providing a unified, end-to-end architecture that tightly integrates denoising (via learnable wavelet shrinkage), multi-scale temporal learning, and advanced graph-based spatio-temporal modeling. I have some comments to further improve the manuscript:

The current work is limited to a small number of sites and climatic regions. While results are promising, the manuscript should more critically discuss generalizability and the steps necessary to adapt WGL to new regions or climates.
The authors should provide a table or analysis of computational complexity, inference latency, and parameter counts for WGL and the baseline models, as deployment in real-time grid environments often faces hardware and latency constraints.
Model robustness to missing/irregular data, calibration drift, and non-stationary weather dynamics should be explored in more detail, either through additional experiments or as a clear direction for future work.
Please proofread for typos and formatting inconsistencies and ensure all references are properly cited and up-to-date.
Please improve clarity and organization, and avoid dense, jargon-heavy paragraphs.

Author Response

Reviewer 3:The paper tackles a pressing challenge in renewable energy integration: robust, accurate, and real-time short-term PV power forecasting at multiple, spatially distributed sites. The proposed WGL framework is innovative, providing a unified, end-to-end architecture that tightly integrates denoising (via learnable wavelet shrinkage), multi-scale temporal learning, and advanced graph-based spatio-temporal modeling. I have some comments to further improve the manuscript:

The current work is limited to a small number of sites and climatic regions. While results are promising, the manuscript should more critically discuss generalizability and the steps necessary to adapt WGL to new regions or climates.
The authors should provide a table or analysis of computational complexity, inference latency, and parameter counts for WGL and the baseline models, as deployment in real-time grid environments often faces hardware and latency constraints.
Model robustness to missing/irregular data, calibration drift, and non-stationary weather dynamics should be explored in more detail, either through additional experiments or as a clear direction for future work.
Please proofread for typos and formatting inconsistencies and ensure all references are properly cited and up-to-date.
Please improve clarity and organization, and avoid dense, jargon-heavy paragraphs.

Response(1):

Thank you for this insightful suggestion. We have added a new subsection in the Discussion that explicitly examines the generalizability of WGL across diverse climatic regions. In particular, we address the potential challenges of transferring the model to different geographical contexts and outline future strategies such as domain adaptation, transfer learning, and the incorporation of regional meteorological priors.

Response(2):

Thank you for the helpful suggestion. We have added “Table 5,” which systematically reports parameter counts (Params), computational cost (MFLOPs), and inference latency (Latency) for all models under a unified hardware and implementation setting. Specifically, WGL has 325K parameters, 5.12 MFLOPs, and a latency of 6.15, achieving the lowest latency among all models while keeping model size and computation moderate; Transformer has 415.75K parameters, 7.275 MFLOPs, and 9.56 latency, and Transformer-LSTM has 415.4K, 6 MFLOPs, and 9.91 latency—despite larger parameter sizes, they maintain relatively low latency due to parallelism; TCN has 139.17K, 2.204 MFLOPs, and 10.11 latency; LSTM has 24.19K, 0.126 MFLOPs, and 11.16 latency; CNN-LSTM has 372.77K, 6.61 MFLOPs, and 14.19 latency; and CNN, although very small (8.129K, 0.086 MFLOPs), exhibits the highest latency at 15.91. These results show that minimal parameters/compute do not guarantee low latency (e.g., CNN is the slowest), Transformer-based models retain a latency advantage through parallelism even with larger sizes, and WGL provides the best overall trade-off among latency, parameters, and computation, making it especially suitable for real-time grid deployment under hardware and latency constraints. The corresponding table and analysis have been incorporated into the manuscript (see Table 5 and the subsequent discussion).

Response(3):

In response to the reviewer’s comment regarding model robustness to missing/irregular data, calibration drift, and non-stationary weather dynamics, we acknowledge that these are critical challenges in practical forecasting systems, especially in the context of photovoltaic power generation. As noted in our conclusion, while the current study demonstrates promising results across various forecasting horizons, several areas indeed require further investigation.First, we agree that robustness to missing data and distributional shifts is a significant area for future work. Although our model performs well under the existing dataset, we intend to extend our experiments to include datasets from a broader range of climatic zones and with more irregularities in the data. This will help assess the model's resilience to missing values and fluctuations in environmental conditions, particularly for PV power generation where weather dynamics can be highly volatile.Furthermore, the model’s performance in the presence of calibration drift—where model parameters may shift over time due to changes in the underlying system or external factors—remains an important open challenge. We plan to investigate adaptive strategies for recalibrating our model as environmental conditions evolve and integrate techniques to mitigate the effects of such drifts, ensuring the model remains accurate over longer operational periods.

Regarding non-stationary weather dynamics, we are aware that the model needs to be robust to varying weather patterns, especially as they impact solar radiation and, consequently, PV power generation. This is an area we intend to explore in more detail in future work. Specifically, we are considering the integration of weather forecasting models and dynamic weather-adjustment mechanisms that can adapt to changing weather patterns, thus improving long-term forecasting reliability.As part of our future research, we will also explore methods for uncertainty quantification, which will allow the model to better handle non-stationary and unpredictable weather scenarios. This will also aid in improving model confidence and robustness in real-world applications.We believe that addressing these aspects will significantly enhance the model’s deployment potential and real-world performance, and we will incorporate these directions in our future research agenda.

Response(4):

Thank you for your valuable feedback. We have carefully proofread the entire manuscript, correcting any typographical errors and formatting inconsistencies. Additionally, we have reviewed and ensured that all references are properly cited and up-to-date. Once again, we appreciate your thorough review, and your feedback has been very helpful in improving the quality of our paper.

Response(5):

Thank you for your feedback. We have revised the manuscript to improve clarity and the overall organization. We have simplified and reorganized some dense, jargon-heavy paragraphs to make the content more accessible and easier to understand. Additionally, we have utilized the language editing service recommended by the journal. We appreciate your suggestion, as it has helped us enhance the readability and coherence of the paper.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

1. “Fig.” should be followed by a space, but not be written as “Fig.1”.

2. The content of Table 4 is overly redundant. It can be optimized to make the presentation more professional and aesthetically pleasing.

3. Please make sure to carefully review the manuscript multiple times to avoid the minor errors.

Author Response

Reviewer 1:

“Fig.” should be followed by a space, but not be written as “Fig.1”.
The content of Table 4 is overly redundant. It can be optimized to make the presentation more professional and aesthetically pleasing.
Please make sure to carefully review the manuscript multiple times to avoid the minor errors.

Response(1)

We thank the reviewer for pointing out this formatting issue. We have thoroughly checked and standardized the entire manuscript (including figure captions, main text, and supplementary materials) to ensure consistent use of the format “Fig. 1,” “Fig. 2,” etc., i.e., with a space between “Fig.” and the number.

Response(2)

Thank you for your valuable suggestion. We agree that the original Table 4 was somewhat redundant and could be improved in terms of clarity and visual presentation. In response, we have carefully revised and optimized Table 4 to enhance its professionalism and aesthetic appeal. The updated version features a more concise layout, clearer structure, and improved formatting to facilitate readability. We believe these adjustments make the table more informative and visually pleasing without compromising any content. The revised Table 4 has been incorporated into the manuscript accordingly.

Response(3)

Thank you for your careful review and kind reminder. We have thoroughly proofread the revised manuscript multiple times to eliminate minor errors and improve overall clarity and consistency. We appreciate your attention to detail, which has helped us enhance the quality of the manuscript.