Accelerating Energy Forecasting with Data Dimensionality Reduction in a Residential Environment

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article explores the method of accelerating energy prediction through data dimensionality reduction in residential environments. By using ANN, 1D-CNN, and LSTM, combined with PCA and VAE, the training time can be effectively reduced.

However, in order to clarify the contribution and improve the readability of the article, some opinions and questions are as follows:

1) This article does not provide a detailed explanation of the specific construction logic of the synthetic dataset (NeurIPS CityLearn Challenge 2023), such as data generation rules, noise introduction methods, and how to ensure its similarity with real residential energy data. Suggest explaining whether the time span of the dataset (92 days) covers seasonal changes (such as winter heating and summer cooling), and if not, discuss its impact on the model's generalization ability.

2) The article mentions that VAE has better nonlinear characteristics, but lacks support for potential spatial visualization or feature importance analysis. Suggest enhancing the credibility of the conclusion by comparing the feature distribution before and after dimensionality reduction (such as t-SNE graph) or reconstructing the error distribution; The training time of LSTM is significantly higher than that of CNN (Table 4), but it has not been analyzed whether it is caused by gradient vanishing/exploding or sequence length limitations, and further discussion is needed.

3) The global model uses 50% of the data from each building, while the local model only uses data from a single building. The difference in data volume between the two may lead to performance bias. It is recommended to supplement the comparison results of the global model under balanced data volume; Figure 10 shows that the MSE of 5-hour prediction is lower than that of 1-hour prediction, but the article does not explain whether this phenomenon is due to the error averaging effect or the long-term dependence of the model. Further analysis should be conducted based on specific cases (such as comparison of predicted curves); The paper mentions that the global model protects privacy, but its security has not been verified through attack experiments (such as inferring user behavior). Relevant discussions or citation support need to be supplemented.

4) The reference citation format is inconsistent in the text, and some references are not labeled according to the standard format, which affects the standardization of the paper. And it is recommended to cite research results from the past five years as much as possible. It is suggested to supplement some of the latest research results to showcase the research progress and achievements in related fields, such as K. Wang et al.'s "Resilience Oriented Two Stage Restoration Considering Coordinated Maintenance and Restructuring in Integrated Power Distribution and Heating Systems" published in 2025.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Reply in the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper presents an approach to energy forecasting using dimensionality reduction, but a more detailed comparison with traditional statistical methods (e.g., autoregressive models) should be added.

Related research for forecasting should be compared: 1: Explainable spatiotemporal multi-task learning for electric vehicle charging demand prediction 2: Multi-node load forecasting based on multi-task learning with modal feature extraction

For the use of PCA and VAE for dimensionality reduction, discuss the potential trade-offs in terms of interpretability and model robustness, especially for real-world applications.

additional statistical significance testing (e.g., confidence intervals or hypothesis testing) should be added for validation of the results.

testing the models on real-world residential energy consumption data would enhance the paper’s contributions.

further elaboration on how different buildings' characteristics (e.g., size, occupancy patterns) affect model performance should be done.

a brief comparison with transformer-based models or other recent advances in time-series forecasting is needed.

a section discussing practical implementation challenges (e.g., deployment in smart meters, privacy concerns) should be added.

Comments on the Quality of English Language

good

Author Response

Reply in the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Good work that presents an analysis of the use of dimensionality reduction in the domain of residential energy consumption, considering three different scenarios for residential energy forecasting: global, individual and detailed by building. Using 3 neural network models: ANN, CNN and LSTM. The results show that dimensionality reduction (VAE) in the CNN model presents the best results in terms of MSE and processing speed.

The methodology, the data used are obtained from a public database, the models and their application and the results are clearly described. The comparison is made in terms of the proposed scenarios and models. The conclusions are adequate. The references are adequate.

Comments
1. Improve the wording of the contribution of the work (lines 43-47), the contributions are not clear.

2. From my point of view, the section "5.3. Selection of the lookback/forecast Windows" and the part about the forecast window should be removed. The comment that a larger forecast window can lead to better results is not generally valid, it only applies to this case.

3. Remove comments from the forecast window from the summary and conclusions.

 

Author Response

Reply in the attached file.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript presents accelerating energy forecasting with reducing data dimension. I have some of major concerns with the manuscript which are listed as follows: 
1. The abstract needs to include clearly how you have addressed the problem. 
2. Please check the literature review, as you have not included significant recent research papers in this field with their limitation that you have solved. 
3. The paper mentions dimensionality reduction, but it does not mention, how this method is different from existing literatures. 
4. Please provide more detail on the model used including proper equations. 
5. The paper lacks comparative analysis with other existing models. 

Author Response

Reply in the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents a well-structured and significant contribution to the field. I recommend acceptance of this manuscript in its current form.

Comments on the Quality of English Language

The language is clear and concise.

Author Response

On behalf of all the authors, thanks again for your feedback.

Reviewer 4 Report

Comments and Suggestions for Authors

All of my concerns are addressed.

Author Response

On behalf of all the authors, thanks again for your feedback.