Energy-Efficient Retrofitting under Incomplete Information: A Data-Driven Approach and Empirical Study of Sweden

Round 1

Reviewer 1 Report

Through performance modeling and data imputation, the authors of this research used Bayesian regularization backpropagation neural networks and fuzzy C-means clustering to facilitate retrofitting selection under imperfect information. The topic is interesting, and the paper itself is well-structured. However, the sections of "discussions" and "conclusions" should be improved.

The findings of this study should be compared with those of previous studies. In the "discussion" section, you should also talk about the possible reasons why your model worked well.

In the "conclusion," the authors should explain the limitations that they faced during their experiments, in addition to the limitations that their model has.

Adding some more quantitative information to the abstract can make it more informative.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This article presents a data-driven approach for the selection of energy efficient measures and retrofit strategies for the building stock in Sweden. The machine learning method aims to overcome the problem of missing data in building performance datasets. The text is interesting and although the method is complex it seems solid to me. I have no major remarks, but the authors should address the comments below to further improve the article: 

1. How did you ensure that the data imputation method (replacement of missing data with substitute value) will not introduce unwanted bias in the data set and affect the data analysis?

2. How flexible is the PMDI method in terms of building climate, building type, construction year, occupants' behavior? How much tweaking and tuning is necessary in the PMDI method to obtain valid results for different building types, in different countries and climates? In lines 387-396 you mention datasets from USA, Japan and Australia - these datasets were also used or only the dataset from Sweden (lines 398-419)?

3. What properties must hold the building performance datasets (BPDs) for the PMDI method to be able to model and analyze the dataset? When will the data cleaning and anomaly detection methods fail?

4. You mention that the data imputation method exploits the correlations between variables to impute case-specific missing data taking into account the relationship between building characteristics and the whole knowledge. Can you give practical example for the imputation process - how to impute the unknown data for external wall U-value, indoor air change rate, or the specific internal heat gains? What type of correlations were used and how were they selected?

5. Table 1 lists the building properties from EPC reports in Sweden. In your opinion, to what extent is the EPC data reliable in Sweden? Among the data items, what is measured by the EPC expert and what is assumed or estimated because it is simpler and faster to do so? You method seems solid to me, but the main flaw could be "garbage in, garbage out" problem. 

6. In Appendix A - Table 1 some of the building contain zero values for "normal year adjusted value (degree days) kWh" and for "the electricity that is included in the building's energy use kWh" - why is that? The largest error in data imputations seems to be the normal year adjusted value (degree days) kWh for building 10 with imputed value of 13571 kWh against real value of 21683 kWh - any particular reason for this large difference?

7. I suggest adding the list of acronyms and symbols.

8. The retrofit measure RM10: timing of ventilation system returns higher energy savings than thermal insulation RM25-27. What would be the explanation for this? Buildings in Sweden are already insulated and adding more insulation does not return large energy savings? What is the meaning of RM9: adjustment of ventilation system?

9. Figure 6-b) shows the annual costs of retrofit measures? RM19 is best for building #5 because of large energy savings and practically zero costs? How can the costs of insulating pipes and ducts be almost zero?

10. Line 634: should be section 4.3. The Conclusions should be section 5.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf