A systematic approach in load disaggregation utilizing a multi-stage classification algorithm for consumer electrical appliances classification

Abstract

The potential to save energy in existing consumer electrical appliances is very high. One of the ways to achieve energy saving and improve energy use awareness is to recognize the energy consumption of individual electrical appliances. To recognize the energy consumption of consumer electrical appliances, the load disaggregation methodology is utilized. Non-intrusive appliance load monitoring (NIALM) is a load disaggregation methodology that disaggregates the sum of power consumption in a single point into the power consumption of individual electrical appliances. In this study, load disaggregation is performed through voltage and current waveform, known as the V-I trajectory. The classification algorithm performs cropping and image pyramid reduction of the V-I trajectory plot template images before utilizing the principal component analysis (PCA) and the k-nearest neighbor (k-NN) algorithm. The novelty of this paper is to establish a systematic approach of load disaggregation through V-I trajectory-based load signature images by utilizing a multi-stage classification algorithm methodology. The contribution of this paper is in utilizing the “k-value,” the number of closest data points to the nearest neighbor, in the k-NN algorithm to be effective in classification of electrical appliances. The results of the multi-stage classification algorithm implementation have been discussed and the idea on future work has also been proposed.

Keywords

load disaggregation / voltage-current (V-I) trajectory / multi-stage classification algorithm / principal component analysis (PCA) / k-nearest neighbor (k-NN)

Introduction

The potential to save energy in existing consumer electrical appliances is very high [¹,²]. One of the ways to achieve energy saving and improve energy use awareness is to recognize the energy consumption of individual electrical appliances [³–⁷]. To recognize the energy consumption of consumer electrical appliances, the non-intrusive appliance load monitoring (NIALM) methodology is utilized [⁷–¹¹]. NIALM is a load disaggregation methodology that disaggregates the sum of power consumption in a single point into the power consumption of individual electrical appliances [¹²]. The energy or power consumption for individual electrical appliance can be determined from the disaggregated data [¹³]. The advantage of using the NIALM methodology is that, there is no need to have physical sensors at each of the electrical appliance monitored.

From the disaggregated individual electrical load, relevant load signature expression can be extracted. Some common load signature measurements are power consumption, transient waveform shape, and harmonics [¹⁴–¹⁹]. In the field of NIALM, there are various researches done that utilize load signatures to identify individual electrical load from cumulative signals [²⁰–²²].

The research conducted by Hart [¹²] utilizes the power differences between ON and OFF states in sequence operations to identify the individual electrical loads. Various researches focusing on other types of load signatures, namely harmonics, transient waveform shape, and energy use pattern of the loads were also used in the electrical load identification [²³,²⁴].

Accordingly, this paper focuses on the multi-stage classification algorithm for image identification of two dimensional voltage-current (V-I) trajectory to classify the individual electrical load. The V-I trajectory for each individual electrical appliance has been plotted using the normalized current and voltage values. Previous work has been done in identifying the shape features of the V-I trajectory with the corresponding individual load and its characteristics [²⁵]. It is shown to be effective in identifying the individual electrical appliance due to different V-I trajectory. Further study has been conducted to the study done by Lam et al. [²⁵] by improving the precision and robustness through the researchers’ algorithm [²⁶]. Hassan et al. have used four different types of learning algorithms, namely the feed-forward artificial neural network (ANN) [²⁷], the hybrid learning algorithm (combination of ANN and evolutionary algorithm (EA)), the support vector method (SVM) with a Gaussian kernel function, and the Adaptive Boost (AdaBoost) algorithm [²⁶]. In their work, the reference energy disaggregation data set (REDD) [²⁸], a publicly available data set was utilized to evaluate the algorithms used and the data sets from REDD’s house # 3 was benchmarked [²⁶].

On top of that, principal component analysis (PCA) is a methodology to reduce the dimension of the data without reducing the critical data in the signature data set [²⁹]. The PCA methodology has been utilized in the face recognition algorithm where the significant features of face images are known as eigenfaces. These eigenfaces are the eigenvectors or principal component of the set of faces [³⁰]. Further researches have been done by various researchers in applying PCA for feature extraction with the classification technique of k-nearest neighbour (k-NN) in the area of fault detection for semiconductor process [³¹] and medical technology [³²,³³]. On top of that, in the classification algorithm, image pyramid reduction methodology is utilized. The pyramid reduction algorithm computes a Gaussian pyramid reduction and reduces the image density and resolution by fixed steps through a low-pass filtered and reduced the image by a factor of two. This is an algorithm to effectively perform image processing in terms of computational cost and complexity [³⁴–³⁷].

Therefore, in the present work, a multi-stage classification algorithm methodology is proposed utilizing the principal component analysis (PCA) and the k-nearest neighbor (k-NN) algorithm that is able to classify the V-I trajectory images for individual electrical appliance. The proposed PCA methodology is an image recognition approach and the k-NN is a supervised learning methodology. The contribution of this paper is in utilizing the “k-value,” the number of closest data points to the nearest neighbor, in the k-NN algorithm to be effective in classification of electrical appliances. To evaluate the proposed classification algorithm and to benchmark with the work done by Hassan et al. [²⁶], the reference energy disaggregation data set (REDD) house # 3 data set [²⁸] have been utilized.

Methodology

The objective of this study is to investigate the feasibility of using the image of V-I trajectory-based load signatures to classify the type of electrical appliances. The k-NN classifier is utilized to perform classification of individual electrical appliance. Figure 1 is an overview of the implementation of the classification algorithm.

Preparation of data set

The REDD data set for House #3 is utilized [²⁸]. This approach is based on the event-based methodology for NIALM [³⁸–⁴¹]. From the REDD data set, the power measurement is obtained [²⁸,⁴¹]. Next, the median filtering is performed to remove noise from the data [⁴¹,⁴²]. After median filtering, the event detection algorithm [⁴¹,⁴³] is performed to identify the changes in the power levels of the appliances from the power consumption profile. The combination of the ON-OFF-based approach [¹²,⁴⁴] and the Goodness-of-Fit (GOF) [³⁸,⁴⁵] techniques are adopted to detect events in the total power consumption data. In House #3, there are 22 electrical outlets that are measured and this includes two mains outlets. The remainder of the 20 outlets correspond to various electrical appliances measured.

Next, from the current and voltage data values (in *.dat format), the data sets are converted to *.mat format. These converted files have 277 columns, in which the first column is the UTC timestamp and the second column is the cycle count. The remaining 275 columns are the waveforms values for the electrical appliances.

Load of residential appliances

The load of residential appliances can be classified into major types, some of which are resistive appliances, pump-operated appliances, electronics appliances, lightings and others [²⁵,⁴⁶].The appliances chosen for this study, based on REDD data set House #3, are as resistive appliances: furnace, washer dryers; pump-operated appliance: refrigerator; others appliances: microwave, bathroom gfi; electronicappliances: lightings.

Table 1 lists the power consumption values and the corresponding channel label for the 10 appliances chosen for the study.

From the event detected waveforms using the GOF methodology from the mains outlet and the timestamp values from the current data for the 10 electrical appliances, a comparison is made. The compared value is the time stamp values. This is to determine that a residential electrical appliance have occurred both in the mains outlet as well as in the current values. The compared and determined timestamp value and waveform are saved in a database.

Next, the comparison algorithm is also applied to the event-detected waveforms using the GOF methodology from the mains outlet and the timestamp values from the voltage data for the 10 electrical appliances, a comparison is made. The compared value is the time stamp values. This is to determine that a residential electrical appliance has occurred both in the mains outlet as well as in the voltage values. The compared and determined timestamp value and waveform are saved in a database.

Figure 2 shows the distribution of power values for the 10 appliances used in House #3 of the REDD data set. Figure 2 also illustrates the extent to which the distribution of power values of appliances overlap with each other. The overlapped power value distributions pose difficulty in discriminating the appliances, especially in cases where the power values fall in the ranges of two or more appliance categories.

V-I trajectory

The timestamp values and waveforms obtained in the comparison in Subsection 2.2 above are being plotted. There are two sets of data obtained, namely, the voltage values and current values from the 10 electrical appliances. For each of the electrical appliances, a normalized voltage versus normalized current, V-I trajectory is plotted. The two figures for each Fig. 3(a) to 3(j) illustrates the two types of typical shapes of the V-I trajectory obtained for each of the 10 electrical appliances.

In this study, 50 V-I trajectory images are prepared as templates train images for each of the electrical appliances. For the test images, 20 V-I trajectory images are prepared as the test images.

Electrical appliances classification algorithm

There are two steps in this classification algorithm of electrical appliances. The first are the crop and image pyramid reduction processes from the original template image and second is the proposed classification algorithm with the combination techniques of principal component analysis (PCA) and the k-nearest neighbor (k-NN) algorithm. Figure 4 is an overview of the implementation of the two-stage classification algorithm.

Template images—crop and image pyramid reduction processes

In this section, the 50 V-I trajectory images are divided into 4 quadrants. Only two quadrants of the cropped image are used, because these are the quadrants which have the signals of most electrical appliances. The quadrants utilized are the top right quadrant and bottom left quadrant. Figure 5 depicts the quadrants. In this work, the top right and bottom left quadrant images are used for load disaggregation because the load of electrical appliances are either resistive or inductive. For capacitive load, the orientation of the V-I trajectory would be different.

After the image is cropped, the size of the image is reduced by half using the image pyramid reduction algorithm. This algorithm is to enable a quicker processing of the subsequent algorithm with the preservation of vital image information. Figures 6 and 7 exhibit an example of the cropped and image pyramid reduction in size by half, the top right and bottom left quadrants.

Classification algorithm

The classification algorithm is a combination of principal component analysis (PCA) and the k-nearest neighbor (k-NN) algorithm.

The PCA is a technique utilizing the statistical method to find the pattern in the data of high dimension. Assuming an image is represented by an N by N dimension, it can be stated as an N₂– dimensional vector [⁴⁷]. An image X can be written as

(1)X=({{\displaystyle x}}_{\mathrm{1}},{{\displaystyle x}}_{\mathrm{2}},{{\displaystyle x}}_{\mathrm{3}}\dots \dots \dots {{\displaystyle x}}_{{{\displaystyle N}}_{\mathrm{2}}}).

The rows of pixels in image X are positioned one after the other to build a one-dimensional image. In the study, 50 images are trained. The few numbers of trained image is based on the previous research that shows the feasibility of training with few examples [⁴⁸]. From the 50 images, each image is N pixels length by N pixels width. For each image, an image vector is created as a one-dimensional image. Next, all these image vectors are place together in one big image matrix as shown in Eq. (2).

(2)\mathrm{Im}\mathrm{age}\mathrm{}\mathit{X},\mathrm{}\mathrm{Matrix=}\left(\begin{array}{c}\hfill \begin{array}{c}\hfill \begin{array}{c}\hfill \mathrm{Im}\mathrm{age}\mathrm{vector\; 1}\hfill \\ \hfill \mathrm{Im}\mathrm{age}\mathrm{vector\; 2}\hfill \end{array}\hfill \\ \hfill ⋮\hfill \end{array}\hfill \\ \hfill \mathrm{Im}\mathrm{age}\mathrm{vector\; 50}\hfill \end{array}\right)\mathrm{.}

Next, the PCA analysis is applied to the image matrix above. The results from the PCA are expressed as principal component coefficients and principal component scores. The principal component coefficients consist of rows of Image X, which represents observations and columns of Image X which represents variables. This coefficient is a P-by-P matrix, where each column contains coefficients for one principal component. The principal component scores are the representation of Image X, and the matrix in the principal component space. The rows of scores correspond to observations and the columns correspond to components.

After the PCA, the train and test scores are placed into the k-NN classifier algorithm. The k-NN classifier finds the nearest neighbor of the data in the training data in the determined feature space by utilizing the Euclidean distance as a distance metric. The majority choice of the nearest neighbor classes will be the class chosen for the data evaluated. The “k” value is the number of closest data points to the nearest neighbor. In this work, the Euclidean distance is used. The k-NN algorithm is chosen as one of the classier in because only one feature is utilized to classify the appliance. This methodology has also been utilized by Gupta et al. [⁴⁹]. In the k-NN classifier, the number of the nearest neighbor, k is varied and different k values are recorded.

In this work, the maximum k-value from all the data points is determined as expressed in Eq. (3). The maximum ‘k’ value for Matlab simulation is determined using Eq. (4).

(3)‘k’\mathrm{m}\mathrm{a}\mathrm{x}=\mathrm{N}\mathrm{o}.\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{d}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{s}\times \mathrm{N}\mathrm{o}.\mathrm{o}\mathrm{f}\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{s}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{p}\mathrm{p}\mathrm{l}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{s}\times \mathrm{N}\mathrm{o}.\mathrm{o}\mathrm{f}\mathrm{s}\mathrm{e}\mathrm{g}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{s},

(4)‘k’\mathrm{s}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{m}\mathrm{a}\mathrm{x}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{m}=\mathrm{0}.\mathrm{5}\times (‘k’\mathrm{m}\mathrm{a}\mathrm{x}).

For example, in the first stage, the ‘k’ max= 10 (No. of trained appliances) × 50 (No. of images for each appliances) × 2 (No. of segments) = 1000. Therefore, the maximum ‘k’ value for Matlab simulation is 0.5 × 1000= 500. The results of simulation for this stage is from k = 1 until k = 500.

In this study, a multi-stage classification is proposed. In the initial stage, all the 10 types of electrical appliances are trained and classified with the corresponding k values, from k = 1 to k = 500, as per Eqs. (3) and (4). In the next stage, the remaining electrical appliances that are not classified in the first stage are placed back into the k-NN classifier. The next stage is also trained and classified. This classification process continues until all the appliances are classified. With the proposed methodology, the proposed algorithm is able to classify the individual electrical appliances studied.

Results and discussion

In the first stage of the classifier algorithm, the 10 selected electrical appliances have been placed into the classifier to classify the corresponding electrical appliances. At this stage, the ‘k’ value for simulation is k = 1 to k = 500. Table 2 lists the result of the first stage of the algorithm.

In the second stage of the algorithm, the remaining electrical appliances, namely, Electronics, Lighting 1, Lighting 2, Lighting 3, Microwave, Refrigerator, and Washer dryer 1 are placed back to the multi-stage classifier for classification. Table 3 is the result of the second stage of the algorithm. Based on this study, in the second stage, at k = 1 to k = 100, the algorithm is able to classify the V-I trajectory image of Electronics (CH06). At k = 110 to k = 120 and k = 170 to k = 180, the V-I trajectory image of Lighting 1 (CH11) is able to be classified by the algorithm. At k = 130 to k = 160, the V-I trajectory image of the Microwave is able to be classified by the algorithm. Next, the V-I trajectory image of Washer dryer 1 is able to be classified by the algorithm at k = 310 to k = 350.

From the results obtained, it is observed that there are overlapping of power values for Lighting 1 (CH11) and Microwave (CH16). With reference to Fig. 2, it is observed that there is an overlapping of power values. Based on this study, the V-I trajectory for Lighting 1 (CH11) and Microwave (CH16) have some distinctive V-I trajectory images to differentiate them.

The next group of V-I trajectory images consist of both Washer dryer 1 and Washer dryer 2. In the first stage of the algorithm, the V-I trajectory image of Washer dryer 2 (CH13) is able to be classified by the algorithm at k = 230 to k = 500 while in the second stage, the V-I trajectory image of Washer dryer 1 (CH14) is able to be classified by the algorithm at k = 310 to k = 350. Both Washer dryers 1 and 2 cover the various modes. This is because based on the REDD data set House #3, there are 537 events detected for WSD 1 (CH14) and 519 events detected for WSD 2 (CH13).

In the third stage, the remaining 2 electrical appliances (Lighting 2 and Lighting 3) that have not been classified are placed back into the classifier algorithm to identify the corresponding appliances. In this stage, the ‘k’ value for simulation is k = 1 to k = 100. Table 4 tabulates the result of the third stage of the algorithm.

Based on this study, the appliances of Lighting 1 (CH11), Lighting 2 (CH17), and Lighting 3 (CH19) have some similarities in the V-I trajectory images and overlapping of power values as indicated in Fig. 2. Therefore, with these overlapping in the images of these 3 electrical appliances, the algorithm is able to classify Lighting 3 (CH19) at k = 1 to k = 17. The remaining k values from k = 8 to k = 100, the algorithm is able to classify Lighting 2 (CH17).

Thus, the proposed multi-stage classification algorithm methodology is able to satisfactorily classify the V-I trajectory images of the 10 electrical appliances. As this is a multi-stage classification algorithm, as shown in Figs. 1 and 4, there is no maximum number of appliance items for disaggregation.

Conclusions

This study was conducted to classify individual electrical appliances from a publicly available data set, REDD, utilizing load disaggregation through V-I trajectory based load signatures. The classification algorithm utilizes a multi-stage classification algorithm methodology. The proposed algorithm is a combination of the principal component analysis (PCA) methodology and the k-nearest neighbor (k-NN) methodology. The classification algorithm is utilizing the image of the V-I trajectory-based load signatures. The novel approach utilizes the PCA-based image recognition method to identify the V-I trajectory image of the electrical appliances with the supervised learning of the k-NN methodology to classify the electrical appliances. The contribution is in utilizing the “k-value,” the number of closest data points to the nearest neighbor, in the k-NN algorithm to be effective in classification of electrical appliances.

Therefore, this research establishes a platform for future research in the direction of utilizing unsupervised learning methodology with the PCA methodology.

References

[1] Crosbie T. Household energy consumption and consumer electronics: the case of television. Energy Policy, 2008, 36(6): 2191–2199

[2] Roth K, Urban B, Shmakova V, Lim B. Residential consumer electronics energy consumption in 2013. In: Proceedings of the ACEEE 2014 Summer Study on Energy Efficiency in Buildings. Washington, 2014, 308–320

[3] Ruzzelli A G, Nicolas C, Schoofs A, O’Hare G M P. Real-time recognition and profiling of appliances through a single electricity sensor. In: Proceedings of the 7th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc Communications and Networks. Boston, USA, 2010, 1–9

[4] Lee S C, Lin G Y, Jih W R, Hsu J Y J. Appliance recognition and unattended appliance detection for energy conservation. In: Proceeding of the Workshops at the 24 AAAI Conference on Artificial Intelligence. Atlanta, GA, USA, 2010, 37–44

[5] Weiss M, Helfenstein A, Mattern F, Staake T. Leveraging smart meter data to recognize home appliances. IEEE Pervasive Computing & Communication, 2012, 22(1): 190–197

[6] Moro J Z, Duarte L F C, Ferreira E C, Dias J A S. A home appliance recognition system using the approach of measuring power consumption and power factor on the electrical panel, based on energy meter ICs. Circuits and Systems, 2013, 4(3): 245–251

[7] Kim Y W, Kong S B, Ko R K, Joo S K. Electrical event identification technique for monitoring home appliance load using load signatures. In: Proceedings of the 2014 IEEE International Conference on Consumer Electronics (ICCE). Las Vegas, NV, USA, 2014, 296–297

[8] Dong M, Meira P C M, Xu W, Chung C Y. Non-intrusive signature extraction for major residential loads. IEEE Transactions on Smart Grid, 2013, 4(3): 1421–1430

[9] Wang Z, Zheng G. Residential appliances identification and monitoring by a nonintrusive method. IEEE Transactions on Smart Grid, 2012, 3(1): 80–92

[10] Streubel R, Yang B. Identification of electrical appliances via analysis of power consumption. In: Proceedings of the 47th International Universities Power Engineering Conference (UPEC). Middlesex, United Kingdom, 2012, 1–6

[11] Figueiredo M, de Almeida A, Ribeiro B. Home electrical signal disaggregation for non-intrusive load monitoring (NILM) systems. Neurocomput, 2012, 96(3): 66–73

[12] Hart G W. Nonintrusive appliance load monitoring. Proceedings of the IEEE, 1992, 80(12): 1870–1891

[13] Chang H H, Lian K L, Su Y C, Lee W J. Power-spectrum-based wavelet transform for nonintrusive demand monitoring and load identification. IEEE Transactions on Industry Applications, 2014, 50(3): 2081–2089

[14] Norford L K, Leeb S B. Non-intrusive electrical load monitoring in commercial buildings based on steady-state and transient load-detection algorithms. Energy and Building, 1996, 24(1): 51–64

[15] Chang H H, Yang H T, Lin C L. Load identification in neural networks for a non-intrusive monitoring of industrial electrical loads. Lecture Notes in Computer Science, 2008, 5236: 664–674

[16] Chang H H, Yang H T. Applying a non-intrusive energy-management system to economic dispatch for a cogeneration system and power utility. Applied Energy, 2009, 86(11): 2335–2343

[17] Chang H H, Chen K L, Tsai Y P, Lee W J. Anew measurement method for power signatures of nonintrusive demand monitoring and load identification. IEEE Transactions on Industry Applications, 2012, 48(2): 764–771

[18] Chang H H. Non-intrusive demand monitoring and load identification for energy management systems based on transient feature analyses. Energies, 2012, 5(12): 4569–4589

[19] Chang H H, Lin C L. A novel information technology of load events detection for the energy management information systems. Information Systems and e-Business Management, 2015, 13(2): 289–308

[20] Laughman C, Lee K, Cox R, Shaw S, Leeb S, Norford L, Armstrong P. Power signature analysis. IEEE Power & Energy Magazine, 2003, 1(2): 56–63

[21] Lee K D. Electrical load information system based on non-intrusive power monitoring. Dissertation for the Doctoral Degree. Boston: Massachusetts Institute of Technology, 2003

[22] Lee W K, Fung G S K, Lam H Y, Chan F H Y, Lucente M. Exploration on load signatures. In: Proceedings of the International Conference on Electrical Engineering (ICEE 2004). Sapporo, Japan, 2014, 690–694

[23] Zeifman M, Roth K. Nonintrusive appliance load monitoring: review and outlook. IEEE Transactions on Consumer Electronics, 2011, 57(1): 76–84

[24] Zoha A, Gluhak A, Imran M A, Rajasegarar S. Non-intrusive load monitoring approaches for disaggregated energy sensing: A survey. Sensors (Basel), 2012, 12(12): 16838–16866

[25] Lam H Y, Fung G S K, Lee W K. A novel method to construct taxonomy of electrical appliances based on load signatures. IEEE Transactions on Consumer Electronics, 2007, 53(2): 653–660

[26] Hassan T, Javed F, Arshad N. An empirical investigation of V-I trajectory based load signatures for non-intrusive load monitoring. IEEE Transactions on Smart Grid, 2014, 5(2): 870–878

[27] Chang H H. Non-intrusive fault identification of power distribution systems in intelligent buildings based on power-spectrum-based wavelet transform. Energy and Building, 2016, 127: 930–941

[28] Kolter J Z, Johnson M J. REDD: A public data set for energy disaggregation research. In: Proceedings of the SustKDD Workshop on Data Mining Applications in Sustainability. San Diego, CA, USA, 2011, 1–6

[29] Jolliffe I T. Principal Component Analysis. New York: Springer, 2002

[30] Turk M, Pentland A. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 1991, 3(1): 71–86

[31] He Q P, Wang J. Principal component based k-nearest-neighbor rule for semiconductor process fault detection. American Control Conference, Seattle. WA, USA, 2008, 1606–1611

[32] Kamath S D, Mahato K K. Principal component analysis (PCA)-based k-nearest neighbor (k-NN) analysis of colonic mucosal tissue fluorescence spectra. Photomedicine and Laser Surgery, 2009, 27(4): 659–668

[33] Mansor M N, Yaacob S, Muthusamy H, Basah S N. PCA-based feature extraction and k-NN algorithm for early jaundice detection. International Journal of Soft Computing and Software Engineering, 2011, 1(1): 25–29

[34] Burt P J, Adelson E H. The laplacian pyramid as a compact image code. IEEE Transactions on Communications, 1983, 31(4): 532–540

[35] Burt P J. Fast filter transforms for image processing. Computer Graphics and Image Processing, 1981, 16(1): 20–51

[36] Adelson E H, Anderson C H, Bergen J R, Burt P J, Ogden J M. Pyramid methods in image processing. RCA Engineer, 1984, 29(6): 33–41

[37] Ogden J M, Adelson E H, Bergen J R, Burt P J. Pyramid-based computer graphics. RCA Engineer, 1985, 30(5): 4–15

[38] Jin Y, Tebekaemi E, Berges M, Soibelman L. Robust adaptive event detection in non-intrusive load monitoring for energy aware smart facilities. In: Proceedings of IEEE International Conference on Acoustics, Speech and Signal Processing. Prague, Czech Republic, 2011, 4340–4343

[39] Anderson K, Berges M, Ocneanu A, Benitez D, Moura J M F. Event detection for non-intrusive load monitoring. In: Proceedings of the 38th Annual Conference on IEEE Industrial Electronics Society (IECON). Montreal, Canada, 2012, 3312–3317

[40] Jazizadeh F, Becerik-Gerber B, Berges M, Soibelman L. An unsupervised hierarchical clustering based heuristic algorithm for facilitated training of electricity consumption disaggregation systems. Advanced Engineering Informatics, 2014, 28(4): 311–326

[41] Yang C C, Soh C S, Yap V V. A systematic approach to ON-OFF event detection and clustering analysis for non-intrusive appliance load monitoring. Frontiers in Energy, 2015, 9(2): 231–237

[42] Arias-Castro E, Donoho D L. Does median filtering truly preserve edges better than linear filtering? Annals of Statistics, 2009, 37(3): 1172–1206

[43] Yang C C, Soh C S, Yap V V. Comparative study of event detection methods for non-intrusive appliance load monitoring. Energy Procedia, 2014, 61: 1840–1843

[44] Giri S, Lai P H, Berges M. Novel techniques for on and off states detection of appliances for power estimation innon-intrusive load monitoring. In: Proceedings of the 30th International Symposium on Automation and Robotics in Construction and Mining (ISARC). Montreal, Canada, 2013, 522–530

[45] Cochran W G. The x² test of goodness of fit. Annals of Mathematical Statistics, 1952, 23(3): 315–345

[46] Sultanem F. Using appliance signatures for monitoring residential loads at meter panel level. IEEE Transactions on Power Delivery, 1991, 6(4): 1380–1385

[47] Smith L. A tutorial on principal components analysis. 2015–05–25, available at otago.ac.nz website

[48] Salperwyck C, Lemaire V. Learning with few examples: an empirical study on leading classifiers. In: Proceedings of the 2011 International Joint Conference on Neural Networks (IJCNN 2011). California, USA, 2011, 1010–1019

[49] Gupta S, Reynolds M S, Patel S N. ElectriSense: single-point sensing using EMI for electrical event detection and classification in the home. In: Proceedings of the 12th ACM International Conference on Ubiquitous Computing. Copenhagen, Denmark, 2010, 139–148

RIGHTS & PERMISSIONS

Higher Education Press and Springer-Verlag GmbH Germany