Power Profiling of Smart Grid Users Using Dynamic Time Warping^†

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a model for analysing power consumption behaviour in smart grids using Dynamic Time Warping (DTW) to classify load patterns and profile consumer usage. The authors demonstrate their approach using real-world data from the AMPds2 dataset, achieving 77.17% accuracy in distinguishing between workday and weekend power consumption patterns. However, a fundamental weakness in this paper is its reductive classification approach that merely distinguishes between weekend and weekday power consumption patterns. This binary classification provides minimal practical utility in the smart grid context and represents a missed opportunity for meaningful analysis. Energy providers and grid operators require far more nuanced consumption insights than simply knowing whether a pattern resembles a weekend or weekday. The 77.17% accuracy achieved for this limited classification hardly justifies the complexity of implementing a DTW-based approach.

The paper fails to establish why this weekend/weekday distinction is valuable for any stakeholder in the energy ecosystem. No compelling use cases are presented for how this classification would improve demand response, energy management, or consumer services. More meaningful classifications could have included identifying specific household activities, detecting anomalous consumption patterns, recognizing specific appliance signatures, or predicting peak demand periods - all of which would offer tangible benefits to utilities and consumers. The authors' claim that this classification helps "forecast power load, detect non-technical loss, and apply appropriate rules to a new consumer" is asserted without demonstration. The paper presents no evidence that a binary weekend/weekday classification enables these outcomes. Furthermore, the dataset spans five months (September 2013 to January 2014), yet seasonal variations in consumption patterns - which typically have far greater significance for energy planning than day-of-week differences - are entirely ignored.

There is disproportionate length devoted to background (Sections 2-3) versus novel contributions, limited experimental validation using only one dataset and simple binary classification, modest performance results without comparison to alternative approaches, and a brief discussion of results without in-depth analysis of failure cases or statistical significance. Several methodological limitations further weaken the contribution, including the use of refrigerator consumption as the sole appliance for profiling, lack of justification for the selected sampling period, and limited exploration of parameter sensitivity. While standard metrics are used for evaluation, the paper lacks cross-validation and provides no comparison with baseline methods or state-of-the-art approaches. Section 5 on privacy implications is valuable but somewhat disconnected from the technical contribution, and mitigation strategies for privacy concerns are not explored. The implementation details of the DTW algorithm are insufficient for reproduction, and the clustering thresholds seem empirically determined without clear methodology.

Overall, the fundamental limitation in problem formulation undermines the paper's contribution regardless of the technical approach used. While DTW is an appropriate technique for time series analysis, a stronger paper would have identified classification targets with clear applications to smart grid management and consumer energy services, not limiting themselves to the refrigerator (one appliance) and for a short timeframe that does not capture seasonality effects.

Comments on the Quality of English Language

Generally ok. Few typos.

Author Response

Comments 1: This paper presents a model for analysing power consumption behaviour in smart grids using Dynamic Time Warping (DTW) to classify load patterns and profile consumer usage. The authors demonstrate their approach using real-world data from the AMPds2 dataset, achieving 77.17% accuracy in distinguishing between workday and weekend power consumption patterns. However, a fundamental weakness in this paper is its reductive classification approach that merely distinguishes between weekend and weekday power consumption patterns. This binary classification provides minimal practical utility in the smart grid context and represents a missed opportunity for meaningful analysis. Energy providers and grid operators require far more nuanced consumption insights than simply knowing whether a pattern resembles a weekend or weekday. The 77.17% accuracy achieved for this limited classification hardly justifies the complexity of implementing a DTW-based approach.

Response 1: Thank you for your insightful feedback. We acknowledge that the current binary classification between workdays and weekends is a preliminary step in our research and does not provide sufficient granularity for smart grid providers. We fully agree that more nuanced consumption insights are necessary for practical applications in energy management. To address this, our future work will refine the profiling approach by introducing more granular categories, such as holidays and other significant time-based distinctions. Additionally, we will improve profiling accuracy by increasing the dataset size, incorporating data from more households, enhancing the sampling rate, and considering a wider range of electrical appliances. Furthermore, we will evaluate our approach using datasets from diverse regions to ensure the generalizability and robustness of our model. These improvements will enhance the utility of power profiling for energy providers and grid operators, making it more applicable to real-world smart grid environments.

Comments 2: The paper fails to establish why this weekend/weekday distinction is valuable for any stakeholder in the energy ecosystem. No compelling use cases are presented for how this classification would improve demand response, energy management, or consumer services. More meaningful classifications could have included identifying specific household activities, detecting anomalous consumption patterns, recognizing specific appliance signatures, or predicting peak demand periods - all of which would offer tangible benefits to utilities and consumers. The authors' claim that this classification helps "forecast power load, detect non-technical loss, and apply appropriate rules to a new consumer" is asserted without demonstration. The paper presents no evidence that a binary weekend/weekday classification enables these outcomes. Furthermore, the dataset spans five months (September 2013 to January 2014), yet seasonal variations in consumption patterns - which typically have far greater significance for energy planning than day-of-week differences - are entirely ignored.

Response 2: Thank you for your insightful feedback. We acknowledge that binary classification between workdays and weekends is a simplified approach that does not fully capture the complexity of energy consumption behaviors. However, this study represents an initial phase in developing a scalable profiling model for smart grid applications. This binary classification serves as a foundational step for testing DTW-based clustering and evaluating its ability to distinguish meaningful load patterns. While we agree that more granular segmentations—such as holidays, seasonal variations, or appliance-level profiling—would provide richer insights, a structured research approach requires validating the model with a simpler classification before increasing complexity.

Additionally, workday and weekend consumption patterns already exhibit statistically significant differences, making them a practical starting point for initial load profiling in smart grids. This provides a baseline for future research, where we will introduce finer categorization to improve accuracy and utility for energy providers. The paper has been updated and this discussion has been added to the second paragraph of subsection 5.3 (Load Data Clustering).

We also incorporated the daily load factor to enhance DTW-based classification, which captures variations in power consumption beyond a simple binary split. As detailed in Section 7 and the Conclusion, our future work will expand this approach by integrating:

• Seasonal variations and holidays
• Higher-frequency consumption data
• Appliance-level signatures and occupancy-based factors

We appreciate your feedback, which aligns with our goal of refining the classification framework to provide more actionable insights for energy management.

Comments 3: There is disproportionate length devoted to background (Sections 2-3) versus novel contributions, limited experimental validation using only one dataset and simple binary classification, modest performance results without comparison to alternative approaches, and a brief discussion of results without in-depth analysis of failure cases or statistical significance. Several methodological limitations further weaken the contribution, including the use of refrigerator consumption as the sole appliance for profiling, lack of justification for the selected sampling period, and limited exploration of parameter sensitivity. While standard metrics are used for evaluation, the paper lacks cross-validation and provides no comparison with baseline methods or state-of-the-art approaches. Section 5 on privacy implications is valuable but somewhat disconnected from the technical contribution, and mitigation strategies for privacy concerns are not explored. The implementation details of the DTW algorithm are insufficient for reproduction, and the clustering thresholds seem empirically determined without clear methodology.

Response 3: Thank you for your detailed and constructive feedback. We appreciate your insights, which will help us refine our work. Below, we address your concerns regarding background length, dataset limitations, experimental validation, methodological aspects, privacy discussion, and implementation details.

1. Background Length vs. Novel Contributions
   1. We acknowledge that Sections 2 and 3 contain extensive background information. However, this was necessary to establish the foundation for our work, given the interdisciplinary nature of the study.
   2. In response to your comment, we have refined these sections to streamline the background discussion and shift the focus toward our contributions. Subsection 4.4 (Smart Grid Data Collection) and Figure 1 were removed from the paper.
2. Experimental Validation and Dataset Limitations
   1. While our study currently relies on a single dataset (AMPds2) and binary classification, this represents an initial phase of our research.
   2. As discussed in Section 7, our future work will incorporate additional datasets from different regions to evaluate generalizability and enhance classification robustness. Furthermore, we will refine our profiling by introducing more granular categories (e.g., holidays and seasonal variations), as well as increasing the number of households and the sampling rate.
3. Performance Evaluation and Benchmarking
   4. We recognize the importance of comparing our results with alternative approaches. We already shown the benefit of DTW’s capability to adapt to temporal variations and compared it with ED for measuring time series distance in subsection 5.2. (Load Data Analysis).
   5. While our current study primarily demonstrates the feasibility of using DTW for load profiling, future iterations will include benchmarking against other time-series clustering techniques such as k-means, hierarchical clustering, and deep learning-based methods.
   6. We also acknowledge that modest accuracy in binary classification may not be sufficient for real-world applications, and we plan to incorporate additional features (e.g., appliance-level profiling, household occupancy data) to improve model performance.
4. Failure Case Analysis and Statistical Significance
   1. We appreciate the need for a more in-depth discussion of failure cases and statistical significance.
   2. In Section 7, we have included a more detailed analysis of classification errors, highlighting edge cases where DTW struggles (e.g., mixed consumption behavior on transitional days like Fridays).
   3. We have also incorporated additional statistical validation to assess the significance of our classification results.
5. Methodological Limitations
   1. Refrigerator Consumption for Profiling:
      1. We agree that relying on refrigerator consumption as the sole appliance may be a limitation. However, this was chosen as a controlled proxy for baseline household activity due to its continuous power draw.
      2. In future work, we will expand our profiling to include additional appliances and test how different appliance combinations affect classification accuracy.
   2. Sampling Period Justification:
      1. The selected sampling period was based on available data granularity in AMPds2, but we acknowledge that further justification is needed.
      2. We plan to explore the impact of different sampling intervals (e.g., 1-minute vs. 10-minute data) and assess their effect on classification performance.
   3. Parameter Sensitivity Exploration:
      1. We acknowledge the importance of a systematic exploration of parameter sensitivity. Future work will include grid search and hyperparameter tuning to optimize DTW parameters.
6. Privacy Implications and Mitigation Strategies
   1. 1. We appreciate your recognition of Section 6 on privacy implications, and we agree that its connection to the technical contribution needs to be clearer.
      2. To address this, we have strengthened the discussion in Section 6, on how load profiling could introduce privacy risks and propose potential mitigation techniques, such as differential privacy and data anonymization. Section 6 has been updated.
7. Implementation Details and Reproducibility
   1. 1. We recognize that the implementation details of DTW clustering were not fully elaborated for reproducibility.
      2. In response, we have updated subsections 5.1 and 5.2 and added a more detailed explanation of the DTW algorithm.
      3. The empirically determined clustering thresholds are now better justified in the revised methodology section.
      4. We will also consider making our code and dataset preprocessing scripts publicly available in future work to enhance reproducibility.

We greatly appreciate the reviewer’s detailed feedback, as it highlights key areas for refinement. While our study represents an early-stage exploration of DTW-based power profiling, we are actively working to enhance its practical value, improve classification accuracy, and strengthen methodological rigor. The revisions in Section 7 now provide a more comprehensive discussion of our study’s limitations and future research directions, and we have refined other sections accordingly to improve clarity and impact. Thank you for your valuable insights.

Comments 4: Overall, the fundamental limitation in problem formulation undermines the paper's contribution regardless of the technical approach used. While DTW is an appropriate technique for time series analysis, a stronger paper would have identified classification targets with clear applications to smart grid management and consumer energy services, not limiting themselves to the refrigerator (one appliance) and for a short timeframe that does not capture seasonality effects.

Response 4: Thank you for your feedback. We acknowledge the concern regarding the problem formulation and the limited scope of our study. Our work focuses on applying DTW for appliance-level energy consumption analysis as a proof of concept, and we selected the refrigerator due to its consistent operation and well-defined consumption patterns. However, we recognize that expanding the classification targets to include multiple appliances and incorporating a longer timeframe to capture seasonality effects would strengthen the practical applicability of our approach in smart grid management and consumer energy services. In Section 7 (Discussion), we discuss these issues and explain our plan to extend the study in future work by:

• Expanding the dataset to include a diverse set of appliances with different operational characteristics.
• Conducting a long-term analysis to capture seasonal variations in energy consumption patterns.
• Exploring additional classification tasks that align more directly with smart grid optimization and demand-side management strategies.

While our current work provides valuable insights into the use of DTW for appliance-level energy profiling, we appreciate the suggestion and agree that broadening the scope would enhance its contribution to smart grid applications.

Reviewer 2 Report

Comments and Suggestions for Authors

1. The introduction is brief, there is room for improvement.
2. The number of data samples affects the accuracy and speed of machine learning. Authors should clarify this.
3. The proposed power profiling model in Figure 2 does not include the detailed algorithm as mentioned in Section 3. Please reconfigure this figure.
4. The accuracy is 77.17%. How can authors improve this figure?
5. The authors should provide more real cases under different situations such as length of day, power load pattern.

Author Response

Comments 1: The introduction is brief, there is room for improvement.

Response 1: Thank you for your feedback. We aimed to keep the introduction concise while ensuring it effectively presents the research motivation, problem statement, and contributions. However, if the reviewer has specific aspects they would like to see expanded—such as more background on power profiling, additional context on smart grid applications, or further discussion of related challenges—we would be happy to refine the section accordingly. We appreciate your insights and look forward to any further guidance.

Comments 2: The number of data samples affects the accuracy and speed of machine learning. Authors should clarify this.

Response 2: Thank you for your insightful comment. We acknowledge that the number of data samples significantly affects both the accuracy and computational efficiency of machine learning models. In our study, we employed DTW for power profiling, which, while effective in handling time-series variations, has a computational complexity that increases with dataset size. To clarify, our dataset was sourced from AMPds2, covering five months of power consumption data, with over 22,000 data samples. While this dataset was sufficient for demonstrating the feasibility of DTW-based profiling, we recognize that expanding the dataset to a longer duration or additional households would enhance model generalization and accuracy. However, this increase in data volume would also impact computational efficiency. To address these challenges, we propose the following optimizations in future work:

• Data Reduction Techniques: Implementing FastDTW or dimensionality reduction methods to maintain accuracy while improving computational efficiency.
• Scalability Enhancements: Leveraging parallel computing and cloud-based processing to handle larger datasets efficiently.
• Adaptive Sampling Strategies: Selecting representative samples dynamically rather than processing all data points, ensuring a balance between speed and accuracy.

We appreciate the reviewer’s suggestion and will further investigate how different dataset sizes influence our profiling model’s performance. Future studies will evaluate trade-offs between data quantity, processing time, and classification accuracy to optimize our approach for real-world smart grid applications. Section 7 (Discussion) has been added to discuss about these.

Comments 3: The proposed power profiling model in Figure 2 does not include the detailed algorithm as mentioned in Section 3. Please reconfigure this figure.

Response 3: Thank you for your feedback. To avoid confusion, we have removed Figure 2 from the manuscript, as it did not fully capture the detailed algorithm outlined in Section 5. Instead, we have ensured that the methodology and steps of the proposed power profiling model are clearly described in the text. If necessary, we can provide an updated figure in future revisions that explicitly aligns with the algorithmic details presented in Section 5. We appreciate the reviewer’s input in improving the clarity of our presentation.

Comments 4: The accuracy is 77.17%. How can authors improve this figure?

Response 4: Thank you for your valuable feedback. The reported accuracy of 77.17% can be improved through several enhancements, as discussed in the Discussion section (Section 7). Specifically, we plan to:

• Increase the Sample Size: Expanding our dataset by incorporating additional real-world power consumption data from multiple households will improve model generalization and classification performance.
• Include More Appliances: Currently, our study focuses on refrigerator power consumption. Extending the analysis to multiple household appliances (e.g., air conditioners, washing machines, and heating systems) will provide richer feature sets for better classification.
• Refine Profiling Segments: Our current classification is binary (workday vs. weekend). Introducing finer segmentation, such as time-of-day variations or seasonal categories, will help capture more detailed consumption behaviors, thereby enhancing accuracy.

These improvements will reduce misclassification rates and increase the robustness of our profiling model, making it more applicable to diverse smart grid environments. We appreciate the reviewer’s insight and will integrate these enhancements into future work.

Comments 5: The authors should provide more real cases under different situations such as length of day, power load pattern.

Response 5: Thank you for your insightful suggestion. We acknowledge that analyzing power consumption under various real-world conditions, such as different day lengths and diverse power load patterns, would enhance the applicability of our profiling model. Currently, our study is based on power usage data from a single household over five months, which provides a strong foundation but has limitations in terms of generalizability. To address this, we plan to expand our study in future work by:

• Incorporating Additional Datasets: Including power consumption data from multiple households across different geographic locations and climate conditions.
• Extending the Analysis Period: Covering at least a full year to account for seasonal variations and longer-term consumption trends.
• Examining Diverse Load Patterns: Profiling multiple appliances and introducing more granular classifications beyond the current binary segmentation (workday vs. weekend).

These enhancements will allow us to evaluate the robustness of our model across different real-world scenarios and improve its practical relevance for smart grid applications. We appreciate the reviewer’s suggestion and will integrate these considerations into our future research.

Reviewer 3 Report

Comments and Suggestions for Authors

This work is interesting and results are well presented. I am quite impressed and therefore proposed minor revision and could be accepted if authors can provide satisfactory replies to my comments in attached pdf.

Comments for author File: Comments.pdf

Author Response

Comments 1: Tables 3 and 4 are distorted. Do improve on it. Is Figure 2 necessary?

Response 1: Thank you for your comments. We agree and accordingly updated the paper. Tables 3 and 4 have been reconstructed and corrected. Figure 2 has been removed from the paper.

Comments 2: There should be a discussion section before the conclusion section to discuss the research findings.

Response 2: Thank you for your suggestion. We have added a new section titled Discussion before the Conclusion section. This section provides an in-depth discussion of the research findings and associated challenges. Additionally, relevant analyses are covered in Subsection 5.6 (Analysis) to further support the findings.

Comments 3: Conclusion section should only be a single paragraph.

Response 3: We agree and have updated the Conclusion Section. It includes only one paragraph.

Comments 4: How does the use of Dynamic Time Warping (DTW) in this study improve power profiling accuracy compared to traditional methods like Euclidean distance?

Response 4: DTW enhances power profiling accuracy by effectively handling temporal misalignments in power consumption time series, a limitation inherent in traditional methods like Euclidean Distance (ED). As discussed in the paper, ED is highly sensitive to phase shifts and primarily focuses on amplitude similarities, meaning even slight time distortions between two otherwise similar time series can result in misleading similarity assessments. Additionally, ED requires time series to be of equal length, posing a challenge when dealing with real-world power consumption data, where missing measurements or variations in sampling rates are common. In contrast, DTW provides a more robust approach by allowing non-linear alignment of time series, making it invariant to time shifts. This ensures that similar consumption patterns are correctly identified even if they occur at slightly different times. As demonstrated in our paper in Fig. 2, ED failed to compute similarity when comparing power usage sequences of different lengths due to missing data, whereas DTW successfully aligned and measured their similarity without issue. By leveraging DTW’s capability to adapt to temporal variations, our model achieves more reliable clustering and profiling of power consumption behaviors, leading to improved accuracy in power profiling. We already updated the reason for using DTW instead of ED in the last paragraph of subsection 5.2.

Comments 5: What are the potential biases introduced by using the AMPds2 dataset, and how might they affect the generalizability of the results?

Response 5: The AMPds2 dataset serves as a valuable resource for analyzing real-world power consumption patterns. However, its use introduces certain biases that may limit the generalizability of our findings. A primary limitation is that AMPds2 contains power usage data from a single household in Canada, which may not adequately represent variations in energy consumption across different geographical regions, climates, household compositions, and lifestyles. These factors, along with cultural habits, appliance types, electricity pricing structures, and seasonal variations, play a significant role in shaping energy consumption behaviors.

Furthermore, the dataset used in this work is confined to a specific time, making it less reflective of long-term consumption trends or emerging patterns influenced by technological advancements, such as the growing adoption of smart appliances and renewable energy sources. To address these limitations and improve the model’s generalizability, our future work will focus on incorporating additional datasets from diverse households across multiple regions. Expanding the dataset to encompass various demographics and energy usage behaviors will provide a more comprehensive assessment of the model’s robustness and adaptability in different smart grid environments.

Comments 6: How does the study address privacy concerns related to power profiling, and what countermeasures could be implemented to enhance consumer data security?

Response 6: Thank you for your comments. In this work, we showed that detecting a user’s refrigerator consumption patterns allows us to distinguish their power consumption behavior on workdays versus weekends. This capability enables the tracking of users’ activities and the profiling of their daily routines. To minimize these issues, our model can be updated by focusing on aggregated consumption patterns rather than individual appliance usage. Differential privacy techniques and secure multiparty computation can further enhance data protection. In addition, anonymization methods can be employed to prevent unauthorized access while maintaining the utility of profiling data. The paper has been updated and this paragraph has been added to Section 6, paragraph 2.

Comments 7: What are the key limitations of the clustering approach used for power load classification, and how could they impact real-world deployment?

Response 7: The key limitations of the clustering approach used for power load classification are:

• Data imbalance: the number of sample power load time series available for training the weekend load pattern set is smaller than that for workdays. This imbalance negatively impacts the clustering accuracy for weekend profiles. This problem can be addressed by incorporating more samples from additional datasets. This fact is mentioned in the last paragraph of subsection 5.6.
• Computational Complexity: basically, the clustering based on DTW algorithm is computationally expensive. The paper acknowledges that DTW requires significant resources when applied to large-scale smart grids and suggests that alternatives like FastDTW or hardware acceleration (e.g., GPU-based processing) could improve scalability. This is discussed in the last paragraph of the Section 7.
• Generalizability Issues: The clustering model is trained on data from a single household (AMPds2 dataset), which may not be representative of diverse smart grid environments with different user behaviors and load patterns. Expanding the dataset to include more households from different regions could enhance the model’s robustness​. We discussed this issue in Section 7.

Comments 8: It’s good to boost up references to > 50 so that its convincing and reliable to the readers. I propose/recommend adding (doi: 10.3390/electronics13030493) to anomaly detection in page 2, line 78 of the manuscript.

Response 8: Thank you for your suggestion. We already added the suggested reference to the paper to improve its reliability.

Comments 9: What are the trade-offs between accuracy and computational efficiency when using DTW for power profiling in large-scale smart grid systems?

Response 9: Thank you for your comments. Due to the possibility of non-linear alignment in time series analysis, DTW provides more precise similarity measurements compared to similar algorithms (e.g., ED). This is particularly beneficial in power profiling, where energy consumption patterns often shift due to variations in user behavior and environmental factors. However, the computational cost of DTW increases significantly with the number of users and the length of the time series, making real-time implementation in smart grid applications challenging. To address this, more scalable approaches are needed. FastDTW [32], a lower-complexity approximation of DTW, offers a viable alternative by reducing computational demands while maintaining accuracy. Additionally, parallelized implementations and hardware acceleration (e.g., GPU-based processing) could further enhance performance in real-time applications. Combining DTW with machine learning techniques, such as clustering or neural networks, could optimize performance by using DTW. We have added this paragraph to the first paragraph of Section 7.

Comment 10: How does the study ensure that power consumption behavior analysis accounts for external influencing factors such as seasonal variations or economic changes?

Response 10: Thank you for your insightful comment. We acknowledge that incorporating additional factors such as seasonal variations, economic changes, and electricity pricing structures can further enhance clustering accuracy and improve the analysis of smart grid users' consumption behavior. In our future work, we plan to generalize our approach by integrating additional datasets from diverse households across multiple regions. Expanding the dataset to encompass various demographics and energy consumption patterns will enable a more comprehensive assessment of the model’s robustness and adaptability in different smart grid environments. This discussion is covered in Section 6 of the paper.

Comments 11: Given the study's focus on workday and weekend load profiling, how might additional segmentation (e.g., holidays, different consumer types) improve prediction accuracy?

Response 11: Thank you for your valuable comment. Incorporating additional segmentation, such as holidays and different consumer types, would further enhance profiling accuracy and provide deeper insights into users’ consumption behavior. Currently, our model captures temporal patterns by distinguishing between workdays and weekends. However, in our future work, we plan to refine this binary profiling approach by introducing more granular categories, such as holidays and other significant time-based distinctions. This expansion will improve the accuracy of smart grid user power profiling by capturing more nuanced variations in energy consumption patterns. This concept is covered in Conclusion Section of the paper.

Reviewer 4 Report

Comments and Suggestions for Authors

The peer-reviewed manuscript concerns a power profiling model for smart grid consumers based on real-time load data obtained from smart meters. In particular, a dynamic time-warping clustering algorithm is used. The manuscript rightly emphasizes that, despite the benefits, the analysis of energy consumption data leads to consumer profiling and creates privacy issues. The manuscript constitutes a solid scientific work and contains key elements of a scientific article. Below are the comments that may help to improve the quality of the final paper:

Comment 1. It is worth considering a separate Research Methodology section, which can be placed after the Introduction section or the Related works section. The methodology should precisely define the research problem, research goals, as well as research methods, techniques, and tools. A general research algorithm can also be presented, e.g. in a diagram.

Comment 2. It is worth indicating the criteria used to select publications for the Related Works section.

Comment 3. The research is based on a solid mathematical foundation. In the mathematical model, it is worth using a key (as in mathematical sciences) under individual formulas, which will help avoid ambiguity in the model.

Comment 4. The statement "We implement DTW using the 297 dtw − python module [68] in Python" (lines 297-298, p. 9) is insufficient. It is worth adding information about what version of the dtw-python package was used, how the package was configured, and other implementation details important from the point of view of computer science.

Comment 5. More details about the empirical tests conducted should be provided. In particular, it is crucial to present the technological stack in terms of the algorithm implementation and testing environment. Hardware and software details should be provided.

Comment 6. It is crucial to precisely indicate the limitations of the presented method.

Comment 7. It is worth considering adding a section Discussion of research results, in which the obtained research findings can be confronted with other studies available in the literature.

Comment 8. In the Conclusions section, it is worth describing in more detail the directions of further research.

Best Regards

Author Response

Comments 1: It is worth considering a separate Research Methodology section, which can be placed after the Introduction section or the Related works section. The methodology should precisely define the research problem, research goals, as well as research methods, techniques, and tools. A general research algorithm can also be presented, e.g. in a diagram.

Response 1: Thank you for your suggestion. We have added a separate Research Methodology section (Section 3) to clearly define the research problem, objectives, methodology, and findings. This section outlines the research methods, techniques, and tools used in our study. Additionally, we have included a general workflow diagram to illustrate the research process.

Comments 2: It is worth indicating the criteria used to select publications for the Related Works section.

Response 2: Thank you for your suggestion. In Section 2 (Related Work), we selected publications based on their relevance to three key domains: power consumption forecasting, anomaly detection, and user profiling. These studies were chosen to ensure a comprehensive review of existing methodologies that align with our research focus. The first paragraph of Section 2 (Related work) has been updated to clarify this selection criterion for better transparency.

Comments 3: The research is based on a solid mathematical foundation. In the mathematical model, it is worth using a key (as in mathematical sciences) under individual formulas, which will help avoid ambiguity in the model.

Response 3: Thank you for your valuable feedback. Our research is indeed built on a solid mathematical foundation, and we appreciate your suggestion regarding the use of a key for individual formulas. To enhance clarity and avoid ambiguity, we have added symbol descriptions and notations to each formula used in the mathematical model. This will ensure that each formula is easily interpretable and aligned with standard practices in mathematical sciences. In Section 5, symbol description and notations are in red color.

Comments 4: The statement "We implement DTW using the dtw − python module [68] in Python" (lines 297-298, p. 9) is insufficient. It is worth adding information about what version of the dtw-python package was used, how the package was configured, and other implementation details important from the point of view of computer science.

Response 4: Thank you for your suggestion. We used the dtw-python PyPI package (version 1.5.3) for our implementation, running it on a Windows 11 machine. The package was used to compute the distance between two time series representing power usage. Our implementation and configuration followed the official documentation available at https://dynamictimewarping.github.io/python/. We have updated Subsection 5.2. (Load Data Analysis) and added more information about dtw-python package, its implementation, and configuration.

Comments 5. More details about the empirical tests conducted should be provided. In particular, it is crucial to present the technological stack in terms of the algorithm implementation and testing environment. Hardware and software details should be provided.

Response 5: Thank you for your valuable feedback. The empirical tests were conducted using power consumption data from the AMPds2 dataset, where measurement details are provided in subsection 5.1 (Data Extraction). The DTW Python package, along with the implementation environment, is described in subsection 5.2 (Load Data Analysis). Additionally, profiling implementation details are outlined in the corresponding section.

Comments 6: It is crucial to precisely indicate the limitations of the presented method.

Response 6: Thank you for your comments. In Section 7 (Discussion), we explicitly address the limitations of our profiling model. These include the constraint of the AMPds2 dataset, which is limited to a single household, and the fact that only one appliance's power usage was analyzed, limiting the model's generalization. Additionally, we discuss the computational limitations of DTW. Furthermore, we outline our future work, which aims to address these limitations and improve the robustness of our approach.

Comments 7: It is worth considering adding a section Discussion of research results, in which the obtained research findings can be confronted with other studies available in the literature.

Response 7: Thank you for your suggestion. A new section, Section 6 (Discussion), has been added to discuss the challenges with this mode. In Subsection 6.1 (Behavioral Insights and Privacy Risks), we discuss privacy issues related to our work and suggest possible solutions to mitigate these concerns. Additionally, since this research is in its initial phase and will be further generalized in future work, we compare our results with similar studies to provide context and highlight the potential contributions of our approach.

Comments 8: In the Conclusions section, it is worth describing in more detail the directions of further research.

Response 8: Thank you for your suggestion. We have already included a detailed discussion on future research directions in the Conclusions section, outlining how we plan to address the identified limitations and expand our study. 

Reviewer 5 Report

Comments and Suggestions for Authors

The paper is of interest, I have some comments to make though:

• How effective is the proposed power profiling model across diverse smart grid environments and consumer behaviors? The paper evaluates the model using data from a single house in Canada. Would the model maintain its accuracy and F-score of 0.8372 when applied to different geographical locations, varying climates, or different demographics with distinct energy consumption habits?

• What is the computational cost and scalability of the Dynamic Time Warping (DTW) algorithm in a large-scale smart grid with numerous users? While DTW is effective for measuring similarity between power usage time series, its computational complexity could be a limitation. How does the model handle the computational demands of DTW when applied to a large number of consumers in a real-time smart grid environment? Are there more computationally efficient alternatives that could provide comparable results?

• How does the model address the trade-off between accuracy and privacy, especially considering the potential for malicious exploitation of user data? The paper acknowledges that power profiling raises privacy concerns. How can the model be adapted to minimize the risk of revealing sensitive user information while maintaining acceptable profiling accuracy? What specific measures are in place to prevent unauthorized access and malicious exploitation of the extracted power consumption patterns?

• What are the limitations of using only refrigerator power consumption data for user profiling, and how does this choice impact the generalizability of the model? The model uses refrigerator power data to monitor consumer power behavior. How would the inclusion of data from other appliances, or a combination of appliances, affect the accuracy and robustness of the profiling results? Is the refrigerator's power consumption pattern universally representative of overall user behavior across different lifestyles and household compositions?

• How does the model account for external factors and contextual information that influence power consumption, such as weather conditions, occupancy patterns, and socioeconomic variables? The paper mentions that a power profile is influenced by consumer behavior and contextual factors. To what extent does the model integrate and account for these external variables to enhance profiling accuracy and provide more comprehensive insights into consumer behavior?

Author Response

Comments 1: How effective is the proposed power profiling model across diverse smart grid environments and consumer behaviors? The paper evaluates the model using data from a single house in Canada. Would the model maintain its accuracy and F-score of 0.8372 when applied to different geographical locations, varying climates, or different demographics with distinct energy consumption habits?

Response 1: Thank you for your comment. The model has been validated using data from a single house in Canada. While this provides strong preliminary results, we acknowledge that variations in geographical location, climate, and household demographics could impact its accuracy. Future work will include testing with datasets from diverse regions to assess generalizability. Additionally, incorporating adaptive learning techniques could allow the model to adjust to different consumer behaviors dynamically​. This plan has been reflected at the the second paragrapgh of the Conclusion section (Section 8).

Comments 2: What is the computational cost and scalability of the Dynamic Time Warping (DTW) algorithm in a large-scale smart grid with numerous users? While DTW is effective for measuring similarity between power usage time series, its computational complexity could be a limitation. How does the model handle the computational demands of DTW when applied to a large number of consumers in a real-time smart grid environment? Are there more computationally efficient alternatives that could provide comparable results?

Response 2: Yes, it is. DTW is computationally intensive, especially when applied to large-scale smart grids with numerous users. While DTW provides robust similarity measurement, we recognize the need for more scalable approaches. FastDTW, a lower-complexity approximation of DTW, could serve as an alternative, offering reduced computational demands while maintaining accuracy​. Parallelized implementations and hardware acceleration (e.g., GPU-based processing) could also improve performance in real-time applications. This issues discussion has been added to Section 7 (Discussion) in the first paragragh.

Comments 3: How does the model address the trade-off between accuracy and privacy, especially considering the potential for malicious exploitation of user data? The paper acknowledges that power profiling raises privacy concerns. How can the model be adapted to minimize the risk of revealing sensitive user information while maintaining acceptable profiling accuracy? What specific measures are in place to prevent unauthorized access and malicious exploitation of the extracted power consumption patterns?

Response 3: Thank you for your comment and concerning for users’ privacy. Our analysis has acknowledged the privacy issues raised by power profiling. To minimize these issues, our model can be updated by focusing on aggregated consumption patterns rather than individual appliance usage. Differential privacy techniques and secure multi-party computation can further enhance data protection​. Additionally, anonymization methods can be employed to prevent unauthorized access while maintaining the utility of profiling data. We have updated Section 6 and this discussion has been added to subsection 6.1, to mitigate possible privacy issues. 

Comments 4: What are the limitations of using only refrigerator power consumption data for user profiling, and how does this choice impact the generalizability of the model? The model uses refrigerator power data to monitor consumer power behavior. How would the inclusion of data from other appliances, or a combination of appliances, affect the accuracy and robustness of the profiling results? Is the refrigerator's power consumption pattern universally representative of overall user behavior across different lifestyles and household compositions?

Response 4: While refrigerator power consumption provides a consistent signal for profiling, we acknowledge that it may not fully capture overall user behavior. Future iterations of our model will incorporate additional appliances, such as TV, HVAC systems and water heaters, to improve profiling robustness​. This multi-appliance approach could help generalize findings across different household compositions and lifestyles. Conclusion section (Section 8) has been updated to indicate our plan for future work.

Comments 5: How does the model account for external factors and contextual information that influence power consumption, such as weather conditions, occupancy patterns, and socioeconomic variables? The paper mentions that a power profile is influenced by consumer behavior and contextual factors. To what extent does the model integrate and account for these external variables to enhance profiling accuracy and provide more comprehensive insights into consumer behavior?

Response 5: Our proposed model currently considers temporal patterns (workdays vs. weekends) but does not explicitly integrate external factors such as weather conditions, occupancy patterns, or socioeconomic variables. Future work will explore incorporating contextual data (e.g., temperature, household size) to refine profiling accuracy​. Machine learning techniques, such as feature selection and regression modeling, could be used to capture these influences systematically. Section 7 (Discussion) has been added to explain the challenges and indicate our plan for future work.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The paper has been improved but the DTW methodology and the limited focus on refrigerator does not make the relationship with the privacy section relevant. Indeed, knowing about refrigerator power draw cannot provide any information about a household's sleeping or shower routines as stated- this is a flaw.

I would still recommend that if authors cannot be bothered to do more temporal analysis, as evidenced by their response, that they select another appliance from the AMPds2 dataset at least. This could be lighting or heating or washing machine, but at least something more meaningful than the periodic refrigerator which has a fairly steady power profile in comparison.

Author Response

Comments: 

The paper has been improved but the DTW methodology and the limited focus on refrigerator does not make the relationship with the privacy section relevant. Indeed, knowing about refrigerator power draw cannot provide any information about a household's sleeping or shower routines as stated- this is a flaw.

I would still recommend that if authors cannot be bothered to do more temporal analysis, as evidenced by their response, that they select another appliance from the AMPds2 dataset at least. This could be lighting or heating or washing machine, but at least something more meaningful than the periodic refrigerator which has a fairly steady power profile in comparison.

Response: We appreciate the reviewer’s detailed feedback and constructive suggestions.

1. Refrigerator Power Consumption and Privacy Concerns:
   We acknowledge the concern regarding the relevance of refrigerator power consumption to privacy discussions. Our intention was not to claim that analyzing refrigerator power usage alone can reveal specific user activities such as sleeping or shower routines. Instead, we demonstrated that distinguishing between workdays and weekends based on refrigerator power consumption is possible. This classification provides insights into user routines at a high level, which can be extended further with additional appliances in future work.

2. Future Work on Temporal Analysis and Additional Appliances:
   While we recognize the value of a broader temporal analysis incorporating multiple appliances, constraints prevent us from extending the analysis in this manuscript. However, we plan to include additional appliances such as lighting, heating, or washing machines in future work to improve classification accuracy and strengthen the privacy discussion.

3. Clarification on Privacy Discussion in Section 6:
   The privacy considerations outlined in Section 6 apply to a general case of smart grid energy consumption analysis rather than being strictly tied to refrigerator power usage. We do not claim that refrigerator consumption alone can infer specific user habits such as showering times. Instead, we highlight the broader privacy risks associated with analyzing energy consumption patterns.

We appreciate the reviewer’s insights and hope this clarification sufficiently addresses the concerns.

Reviewer 2 Report

Comments and Suggestions for Authors

The revised manuscript is recommended for publication in its current form.

Author Response

Comments: The revised manuscript is recommended for publication in its current form.

Response: Thank you for your time and valuable comments to improve the quality of our manuscript.

Reviewer 5 Report

Comments and Suggestions for Authors

no further comments thanks

Author Response

Comments: no further comments thanks

Response: Thank you for your time and valuable comments to improve our paper's quality.

Round 3

Reviewer 1 Report

Comments and Suggestions for Authors

I am afraid that authors have not adequately addressed my comments. While the methodology, results and conclusions are sound, the novelty is only very marginal. In the interest of scientific rigour, the authors must embed the future work they adeptly stated into the manuscript before it is suitable for accept.