Noninvasive Detection of Appliance Utilization Patterns in Residential Electricity Demand^†

Round 1

Reviewer 1 Report

Interesting paper, but under my opinion have a lacks that is neccesary to improve.

1.- The study is to short. Is neccesary almost 1 year to do a realiable study and good results.

2.- 6 and 3 seconds is a little granularity time. Author can found studies that show the importance of this time. With this granularity time the loss of information is very important, Authors can find papers about this in https://www.nature.com/articles/sdata2016122  https://www.sciencedirect.com/science/article/abs/pii/S0960148115304675?via%3Dihub

http://dx.doi.org/10.1109/PTC.2013.6652217

https://www.mdpi.com/1424-8220/20/21/6034

This is only a few examples. Please check an add this granularity in Introduction. And then incorporate the influences in the methodology of the paper.

2.- Why use UK-DALE and REDD database. Also ara available otrhe more complet and with low granularity resolutions and long register period up than 1 year. By example https://www.pecanstreet.org.

 

Major revision.

Author Response

1.- The study is to short. Is neccesary almost 1 year to do a realiable study and good results.

ANSWER: We thank the reviewer for the comment. While we partly agree with the statement, a study of one year would lead to variations related to seasonal changes that might lead to unsound results. Our main objective is to propose a general methodology that could be employed in other datasets and in practical setups in a straightforward manner, also considering possible variations that would require new training periods. For example, the COVID situation led to significant variations in electricity demand in households, and the proposed methodology could be directly to this. In this sense, we believe that a one year long dataset is an interesting topic for investigation but is beyond the scope of the conceptual apparatus presented here.

However, we agree with the reviewer that this issue was unclear in the previous version of the manuscript. Hence, we added the following paragraph in the concluding section of paper (Page 21, marked in bold): 

In summary, our main objective was to propose a general methodology that can be employed to other datasets and practical setups in a straightforward manner, while also considering possible variations that would require new training periods. In a future paper, we will attempt to generalize this methodology further by applying it to yearly data as offered by other datasets [33,34], especially taking into account seasonal and other variations.

######

2.- 6 and 3 seconds is a little granularity time. Author can found studies that show the importance of this time. With this granularity time the loss of information is very important, Authors can find papers about this in https://www.nature.com/articles/sdata2016122  https://www.sciencedirect.com/science/article/abs/pii/S0960148115304675?via%3Dihub

http://dx.doi.org/10.1109/PTC.2013.6652217

https://www.mdpi.com/1424-8220/20/21/6034

This is only a few examples. Please check an add this granularity in Introduction. And then incorporate the influences in the methodology of the paper.

ANSWER: We appreciate this comment that the granularity of the data is affecting the outcome of the algorithm. The rationale to keep the granularity is the following: (i) it allows to capture the activity at lower levels and then provide the “signature” of different appliances and (ii) it was the granularity available for the most of suitable datasets available. We would like to reinforce that the granularity will probably change the specific numerical results, but the proposed methodology will work as well. The main question is how the methodology would be used by the network operator. It is also important to remark that what is needed for the proposed methodology to work is that the measurements are synchronous, which allows the proposed data processing to work for both PCA and SOM methods.

To explicitly include the aspects related to the granularity, we have added the following paragraph with the references suggested by the reviewer in page 3 (marked in bold).

REDD and UK-DALE datasets were selected due to their consistency and low granularity that enable the appliance status to be captured with a high degree of precision.  Details about  the  impacts  of  the  time  granularity  in  electricity  metering  can  be  found  in  [23–25].   However,  as we are going to discuss later,  the proposed analysis is also valid for other time granularities; the only necessary condition is the time synchronization of the time series.

In page 6 (lines 143-146), we wrote:

It is also worth mentioning that the proposed methodology does work with different time granularities, considering that it allows for detecting the activity of the different appliances. For this, their individual time series must be synchronized to be able to find the patterns of joint usage.

####

2.- Why use UK-DALE and REDD database. Also ara available otrhe more complet and with low granularity resolutions and long register period up than 1 year. By example https://www.pecanstreet.org.

ANSWER: Many thanks for this suggestion. We were actually aware of Pecan Street when we started our research. However, at that time (early to mid 2020), they were modifying their access policies, not making the dataset freely available. It seems that this is not the case nowadays (it wasn’t in 2018 as well, when some of the co-authors have worked with it). Testing our approach in Pecan Street would be an important next step for our research; however, at this point, we believe that this is unfeasible because of the strict deadline for resubmission during this Major Revision. Nevertheless, as argued above, we think that our main contribution is solid without this additional test. For the reasons cited above, we also think that REDD and UK-DALE are the most suitable datasets due to its consistency and low granularity.

Please, refer to the new parts included based on your previous comments

Reviewer 2 Report

• Chapter 2.2.1 “Principal Component Analysis” is unnecessary detail and does not represent the scientific contribution of the paper than the already known method, the chapter could be shortened

• The presentation of Figure 4 should be improved

   

Author Response

Chapter 2.2.1 “Principal Component Analysis” is unnecessary detail and does not represent the scientific contribution of the paper than the already known method, the chapter could be shortened

The presentation of Figure 4 should be improved

ANSWER: Thanks for the suggestions, we have amended the text and Figure 4 accordingly. Please refer to the new Sec. 2.2.1, where the details of the PCA have been removed.

Reviewer 3 Report

a) The authors claim that: "Using the proposed clustering techniques, system operators can implement effective demand-side management". How is that properly justified by this research? Demand-side management with which groups/clusters of appliances?

b) The auhors also claim that: "the system status classifier can be used to detect appliance malfunctions through system status analyses alone". How is that possible and how is that proved with this work? Appliance malfunction cannot be detected with low-freq data and that's why companies that do that usually work with kHz or even MHz of data resolutions. What do the authors have to comment on that?

c) The solution proposed is indeed noninvasive by means that no information is required by the households. But on the other hand if there is not an (accurate) disaggregation block as part of this process then this is in total invasive and intrusive since it relies on the plugs information. What do the authors comment here?

d) Something I am missing and usually is a must have for journal publications is a benchmark analysis against other pre-published literature. The authors need to provide some kind of proof that the proposed implementation is indeed unique and compared to others is better, more accurate etc.

e) I recommend also to further enhance the references list with more and recent papers published on this topic.

Author Response

a) The authors claim that: "Using the proposed clustering techniques, system operators can implement effective demand-side management". How is that properly justified by this research? Demand-side management with which groups/clusters of appliances?

ANSWER: Studying the loads usage patterns in industrial or commercial buildings may be an easy and even obvious work. It may be easy in industries because the process management system reads the system status and commands it (turning on or off pumps, lights, cooling systems), and because the industrial process demands the sequence of system status, with little modulation possibilities. It may be obvious in commercial buildings because the most important loads (in terms of power consumption) are well known (air conditioning, projectors, printers), also with little modulation possibilities because of the commercial hours. 

On the other hand, in residential installations, first each family has its own “process”, which means the operation routine regarding the house appliances. Also, the main loads are not obvious and, although there are many modulation possibilities, the change of habits is not easy to be implemented without guidance. In other words, regular people need guidance/suggestions to improve the house energy efficiency. The clusters’ detection opens the possibility to the energy suppliers to provide (if requested by the final consumer) some personalized suggestions regarding changes of habits that can help improve the residence energy efficiency.     

We added our motivation in page 2 (marked in bold).

Our motivation for studying the loads usage patterns comes from industrial or commercial buildings where the process management system reads the system status and commands it (turning on or off devices such as pumps, lights, and cooling system), as indicated in [20-22], but with the aim to extend it to households in a non-intrusive manner.

###

b) The auhors also claim that: "the system status classifier can be used to detect appliance malfunctions through system status analyses alone". How is that possible and how is that proved with this work? Appliance malfunction cannot be detected with low-freq data and that's why companies that do that usually work with kHz or even MHz of data resolutions. What do the authors have to comment on that?

ANSWER: The SOM, when used as a classifier, can tell us if the system status is, first of all, usual or not usual, in statistical terms. For example, the pair “desktop + screen” are usually used together. A status with desktop “on” and monitor “off” is at least unusual, but can represent a malfunction of the screen. The opposite situation can mean that the desktop failed in turning off, or that the user forgot to turn it off. The method does not intend to detect malfunctions as internal short circuits, but rather problems are detected by the fact that the system status is statistically unexpected. We agree with the reviewer that this point should be clarified in our manuscript, thus we have amended in page 21 (lines 434-436), to include a discussion about this issue.   

(...) the system status classifier can be used to detect appliance malfunctions by indicating system statuses that are statistically uncommon. Note, though, that the method does not intend to detect malfunctions as internal short circuits.

#####

c) The solution proposed is indeed noninvasive by means that no information is required by the households. But on the other hand if there is not an (accurate) disaggregation block as part of this process then this is in total invasive and intrusive since it relies on the plugs information. What do the authors comment here?

ANSWER: Many thanks for this comment and we agree with the reviewer. This was indeed discussed among the co-authors during this study. The idea is that there should be an accurate disaggregation block, and not the information of the plugs (which is both intrusive and unfeasible).

To clarify this item, we had added the following in page 6 (lines 140-143).

Note that the method is developed utilizing disaggregated data. If plugs are directly used for obtaining this data, then the method is clearly intrusive and unfeasible to become a scalable solution. However, we anticipate that in more realistic scenarios, an additional block will be included for data disaggregation, which accurately maps the utilization of the appliances (e.g., [28]).

###

d) Something I am missing and usually is a must have for journal publications is a benchmark analysis against other pre-published literature. The authors need to provide some kind of proof that the proposed implementation is indeed unique and compared to others is better, more accurate etc.

ANSWER: While we generally agree with the reviewer, in this specific case, the authors are not aware of any benchmarks with which this proposal could be directly compared. To the best of our knowledge, this method is novel in comparison to other noninvasive methods that focus on the detection of appliances, not the patterns of joint consumption. In fact, the methodology to perform this is the main contribution of the present paper. If the reviewer is aware of any other method that our proposal could be compared to, we would be very glad to implement and carry out such an important comparison.

###

e) I recommend also to further enhance the references list with more and recent papers published on this topic.

ANSWER: We thank the reviewer for drawing our attention to this issue. We agree with the reviewer, thus we have added more, recent references in the paper, also indicating other aspects that are relevant to the proposed method as, for example,  granularity of the data. Please refer to the parts of the paper marked in bold. New references are: [20]-[25], [28], [30]-[34].

Round 2

Reviewer 1 Report

Accept

Reviewer 3 Report

thanks for your responses, no further comments