A Step-By-Step Identification Method For Non-Intrusive Loads Based On Fused Features - Removed
A Step-By-Step Identification Method For Non-Intrusive Loads Based On Fused Features - Removed
A Step-By-Step Identification Method For Non-Intrusive Loads Based On Fused Features - Removed
Abstract: It is likely to fail to correctly identify loads with similar voltage-current (V-I) trajectories
within the non-intrusive load recognition method, besides the training parameters of the recognition
model are too many by using conventional deep learning algorithm, and it also uses up more
computer resources. Therefore, the use of fused features and a step-by-step identification algorithm
are provided for non-intrusive loads combining the Technique for Order Preference by Similarity to
the Ideal Solution (TOPSIS) algorithm and the Inception network in this paper. Firstly, the Inception
network is applied to recognize the loads with V-I trajectories in the first step; then the loads that
are not successfully identified in the first step can be identified using fused numerical features in
the second step based on the TOPSIS algorithm to achieve fast and fine recognition results. The
proposed method reduces the computational cost and enhances the performance of load recognition,
according to the validation and comparative analysis with other algorithms on the Plug-Load
1. Introduction
According to China's dual carbon target, the refined management of the electricity demand side
plays a significant role in the rational dispatch and effective use of electrical energy, and it has led
to continuous innovation of the power system [1]. Non-intrusive Load Monitoring (NILM) is the
key foundation of power management on the demand side. Customers can regulate their electricity
consumption behavior to save electricity by using NILM technology, which also allows the grid to
optimize the system's structure and increases the efficiency of the power dispatch. Consequently,
the study of NILM with a highly practical significance has attracted scholars’ attention [2][3].
A conventional method of non-intrusive load recognition was based on dividing load features
into steady-state and transient features, and usually combined with typical classification algorithms
such as deep learning for identification. A conventional method based on power features was
presented in [4] compared with a classification method based on V-I trajectories, and the results
showed that the classification method achieved a better recognition rate than the conventional
method. Based on V-I trajectories with the amplitude transformed to pixel values, a method for
distinguishing various kinds of loads was presented in [5] based on a two-dimensional (2D)
convolutional neural network (CNN). [6] further introduced color-coded V-I trajectories and used
pre-trained CNN to distinguish loads in the PLAID dataset. Both color-coded and pixelated V-I
trajectories have a higher recognition rate than conventional V-I trajectories, but there is a
disadvantage of taking up computer resources in these methods. Then, [7] first mapped V-I
trajectories onto cells with binary values and used a Supervised Self-Organizing Map (SSOM) for
recognition. This method greatly decreased the computational work compared to other conventional
methods, and the results showed that the recognition rate was higher than the rate in [6].
However, only using V-I trajectories is unable to distinguish various loads with similar V-I
trajectories and to accurately reflect the other features of loads, such as harmonics.
To address this drawback, boosting features carry by loads can effectively improve the
recognition rate based on V-I trajectories. Fused color V-I images with current, voltage, and phase
were presented in [8], and the method realized the effective identification of different types of loads.
[9] extracted V-I trajectories and temporal features for recognition based on CNN and the Long and
Short Term Memory (LSTM) architecture, and the results showed that the model outperformed the
remaining load identification method on different loads. [10] and [11] combined CNN with a deep
fusion of harmonics, power, and V-I trajectories in a high-dimensional environment, and the results
showed that this method was also effective in identifying different types of loads. [12] and [13] used
back propagation (BP) combined with CNN to form a two-channel model to identify loads, and the
results showed that the identification accuracy was significantly improved for the loads with multi-
states.
Even though the deep learning method of fused features in combination with dual channels
enhanced the extraction of various features, its training process necessitates a significant amount of
computing, and it was a heavy burden on the computer's processing capability. Conventional CNN
usually used deeper convolution to extract high-dimensional information, however, they had an
excessive number of training parameters, which was prone to overfitting phenomena. The Inception
network is a particular kind of CNN, its network structure can be expanded horizontally, and the
input feature vectors can be extracted at various scales, it can increase the number of neurons in the
network, and significantly reduce the number of network parameters [14]. A network model
combining the Visual Geometry Group (VGG) and Inception network in the field of image
processing was proposed in [15], and the results showed that the use of the Inception network can
Additionally, the algorithm with relatively low computational complexity is also a current
research direction for load recognition. [16] identified loads with similar current features using a
similarity matching algorithm based on TOPSIS, but the recognition rate was insufficient since there
was no feature fusion. To improve the recognition rate, the fused features and a step-by-step
recognition method are used to identify loads with similar V-I trajectories based on Inception-
1. It is the first time to apply the Inception network to research non-intrusive loads
identification. And the calculation cost can be greatly reduced by using the advantages of
2. The step-by-step recognition method can effectively improve the recognition rate based on
multi-features.
The paper is organized as follows: Section 2 introduces the algorithm's flow and the step-by-
step recognition method. Section 3 presents the recognition principle of the first step based on the
Inception network. Section 4 describes the recognized principles of the second step based on the
TOPSIS algorithm. Section 5 identifies 11 classes of loads in the PLAID dataset and demonstrates
the effectiveness and accuracy of the proposed algorithm. And the algorithms in this paper are also
compared and analysed with other algorithms. Finally, Section 6 concludes this study.
2. Principles of Step-by-Step Non-intrusive Load Identification based
Fig. 1 shows the step-by-step recognition algorithm of the non-intrusive loads. The first step is
to recognize the V-I trajectories by using the Inception network, and the second step is to recognize
the fused numerical features of loads that are unidentified in the first step by using the TOPSIS
algorithm, and the identification results are obtained by these two steps.
Usually, the V-I trajectory contains rich steady-state features and loads can be distinguished by
The binary-mapped 2D V-I trajectory is a mapping of the V-I trajectory onto cells, each cell is
labeled with a binary number [17]. The matrix cell will be set as 1, if a trajectory passes through the
element of the im -th row and vm -th column of the matrix, otherwise it is 0. And the V-I values in
the steady-state are linear, they can be converted to integers between 0 and n using Equation (1),
I m min I
im n m 1, 2,3, ,n
max I min I
(1)
Vm min V
vm n m 1, 2,3, ,n
max V min V
Where: im , vm are the converted current and voltage at the m -th sample point, respectively.
I m , Vm are the raw current and voltage of the data point m , respectively.
min I , minV are the minimum values of current and voltage in a steady-state period,
respectively.
max I , maxV are the maximum values of current and voltage in a steady-state period,
respectively.
The 2D V-I trajectories of loads in the PLAID dataset are shown in Fig. 2. Some loads with
similar V-I trajectories cannot be effectively distinguished, while others with different V-I
3.2 First Step Identification by the Inception Network based on V-I Trajectories
The Inception module is shown in Fig. 3. It performs different convolution and pooling
operations such as 1×1, 3×3, and 5×5 on the input data. Unlike the conventional sequential
concatenation of convolution and pooling layers, the Inception module cleverly achieves
dimensionality reduction and allows simultaneous convolution and aggregation at multiple
dimensions by using 1×1 convolutional substitution. The output of each part is pooled by using the
same type of padding, which fills the dimensions of the output with the same as the input. Finally,
Global Average Pooling (GAP) layer, as depicted in Fig. 4. One Inception module is noted as one
Block, and the convolution and pooling steps of Block 1 are set to 2, the convolution and pooling
steps of Block 2 are set to 1, and the activation function is set to "Relu". A 2D V-I trajectory is
passed different convolution and pooling operations, and it is subsequently output to the Softmax
function through a GAP layer, then the final recognition result will be output.
a Softmax function to classify data. Due to the huge number of fully connected layers, the training
speed of the model is slowed down, it is also likely to cause an overfitting phenomenon. Each feature
map of the convolution output is averaged by replacing the traditional fully connected layer with a
GAP layer, and the need for training parameters is eliminated, thus the network parameters are
drastically reduced and resulting in a more robust model with the reduced overfitting phenomenon
[18]. Fig. 5 shows a comparison of the fully connected layer and the GAP layer. It is easy to see that
the parameter calculation of the Inception network has greatly reduced by using the GAP layer
instead of the fully connected layer from this simple comparative structure diagram alone.
Fig. 5. Comparison between the Fully Connected Layer and the GAP Layer
The V-I trajectories alone are not sufficient to effectively differentiate loads with similar
trajectories in the first step, so the second step is required to achieve the recognition of the surplus
loads. The fused numerical features, which combines active and reactive power, harmonics, and
power factor [19], will be used for load recognition based on the TOPSIS algorithm.
Different numerical features can effectively distinguishs different loads. In this paper, the
active and reactive power, the first five odd harmonics, and the power factor are fused as numerical
features. Among them, active and reactive power features can distinguish between high and low
power loads, harmonics can compensate for the problem on the lack of sampling accuracy with the
V-I trajectory, and power factors can reflect the energy efficiency of loads and distinguish the nature
of loads. The fused numerical features are extracted by using the Fast Fourier Transform (FFT) [20].
As Equation (2), X (k ) is the k -th signal of the frequency domain after the FFT during a
specific steady-state:
N 1 2 πkn
j
X (k ) in e N
, k 0,1,, N 1 (2)
n0
Where, in is the i -th current sampling. N is the total number of sampling points in a period.
X (k ) is the amplitude of the k -th signal, it is also the k -th harmonics. However, the odd
harmonics of household loads are significantly larger than the even harmonics, and the harmonic
value will be already small when the order of harmonics is greater than 11 [10]. Therefore, the first
Additionally, the time domain voltage and current can be expressed by Equation (3):
N 1
i(t ) I (0) 2I ( k ) cos kwt φi( k )
k 1
N 1
(3)
v(t ) V (0)
2V (k )
cos kwt φ (k )
v
k 1
I (0) , V (0) are the rms fundamental current and voltage, respectively.
I ( k ) , V ( k ) are the rms values of k -th harmonic current and voltage, respectively.
i( k ) , v are the phase angles of the k -th harmonic current and voltage, respectively.
(k )
Then, the active and reactive power can be calculated by the time domain voltage and current
by Equation (4):
N 1
P V (0) I (0) V ( k ) I ( k ) cos φv( k ) φi( k )
k 1
N 1
(4)
Q V (k ) (k )
I sin φ (k )
v φ
(k )
i
k 1
While the power factor is generally expressed as λ and calculated from the active and reactive
P
(5)
P Q2
2
In this paper, the evaluation index refers to the numerical features of each load in the database,
and the evaluation object refers to each load that was not successfully identified in the first step. Fig.
which combines the entropy weight method [22], the Criteria Importance Though Intercriteria
Correlation (CRITIC) weight method [23], and the Principal Component Analysis (PCA) weight
method. These three objective weight methods can effectively prevent errors caused by subjective
factors and ensure the integrity of the evaluation process. The entropy weight method
emphasizes the variability of features and reflects the differences between features, the CRITIC
fully considers both the correlation and the conflict between features, whereas PCA focuses on the
correlation between features. These three weight methods are combined to obtain the comprehensive
W j1W j2W j3
Wj M
(6)
W W W
j 1
1
j j
2 3
j
Where M indicates the number of load features, W j1 , W j2 and W j3 denotes the entropy weight,
Finally, the identification results of the second step are obtained based on the similarity Di
between the load to be tested and the i -th load in the database. The greater the Di is, the more
similar the load to be tested is to the loads in the database, then the tested load can be identified and
The PLAID dataset [24] has a total of 1074 datasets from 235 independent loads and provides
voltage and current for 11 different types of household loads at a sampling frequency of 30 kHz for
experimental testing.
The average of the F1-score in the first step and the accuracy ε in the second step for each load
is the accuracy for the final recognition rate of the step-by-step identification model.
There are three evaluation indicators for identification in the first step. Accuracy, F1-score, and
confusion matrix [25] are used to evaluate the recognition rate for the Inception network. Where
Accuracy is used to evaluate the performance of the model, and the F1-score is used to evaluate
whether each load is identified successfully in the first step of identification. Among them, Accuracy
And the Precision is the proportion of all samples with accurate positive predictions out of all
TN indicates the number of classes that are negative but are predicted to be negative.
FP indicates the number of classes that are negative but are predicted to be positive.
FN indicates the number of classes that are positive but are predicted to be negative
[26].
Finally, the F1-score is the summed average of the Precision and Recall, it can be obtained by
Equation (10).
2Pre Rec
F1-score (10)
Pre Rec
The accuracy indicator ε for the second step represents the ratio of the number of correctly
identified loads nture to the total number of loads ntext , it can be calculated by Equation (11):
nture
ε= 100% (11)
ntest
Firstly, this paper constructed the 2D V-I trajectories image. And the number n was set to 32.
After adjusting the relevant parameters, the ratio of the training set and the test set is 8:2, and
the training process is shown in Fig. 7. The recognition accuracy and loss value tend to be stable
when the epoch number exceeds 60, and the difference of the accuracy between the test set and the
training set is minimal. It shows that the Inception network has a good anti-overfitting ability.
shown in Fig. 8. The accuracy for the Inception network model achieves 96.26%.
Fig. 8. Confusion Matrix for the Recognition with Inception Network
(Where AC: air conditioner, CFL: compact fluorescent lamp, Bulb: incandescent light bulb,
Washing: washing machine)
The F1-score for each type of load, which is derived from the confusion matrix, is used to
assess the model's capability to recognize loads. As shown in Fig. 9, the F1-score of the Fridge, AC,
and Heater are all lower than 90%, because these loads have multiple operating states, and the
features of similar loads in different states are quite different. The CFL and Laptop both achieves
an F1-score of 1, and the recognition accuracy is 100%. Except for these two loads, the remaining
loads are not correctly recognized, and they will be recognized in the second step.
In this subsection, the numerical features of the loads except for CFL and Laptop are acquired,
and the feature database is built. A total of 180 loads are randomly selected from 9 types, and the
recognition results for each load will be obtained by calculating the similarity using the TOPSIS
algorithm, and the recognition results are shown in Table 1.
ε (%) 98.12 99.07 98.56 97.23 100 100 100 100 96.57
The numerical features of the Fan and CFL are more similar in terms of power factor and power
distribution. And the identification rate of the Fan will be effectively increased when the CFL does
Then, an average recognition rate of 99.12% can be attained for the step-by-step recognition
Performance comparisons between the conventional CNN and the Inception network were
made in this subsection. The CNN consists of three convolutional and pooling layers, which are
followed by a fully connected layer, and then it uses Softmax for classification. The training process
is shown in Fig. 10. The recognition accuracy and loss value also tends to be stable when the epoch
number exceeds 60, and the difference in the accuracy in the training set between the Inception and
conventional CNN is about 0.1. And the loss value of Inception is significantly lower than CNN's.
Therefore, the Inception network has higher recognition accuracy, faster convergence, and smaller
Fig. 10. Training Process Comparison of Inception Network and Conventional CNN
As shown in Fig. 11, the confusion matrix of the conventional CNN can achieve a recognition
accuracy of 85.51%, which is 11 percentage points lower than the Inception network compared with
Fig. 8. However, none of the F1-scores of the loads reached 1 as shown in Fig. 12. Therefore, all 11
types of loads were not successfully identified in the first step and need to be identified in the second
step.
Fig. 11. Confusion Matrix for the Recognition with Conventional CNN
shown in Table 2. The highest accuracy ε is 98.09 %, and it is significantly lower than the accuracy
And the average accuracy of the step-by-step recognition based on conventional CNN and
TOPSIS is 97.02%, which is two percentage points lower than the proposed method in this paper.
Obviously, based on the analysis above, the step-by-step recognition by using the Inception-TOPSIS
model improves the recognition accuracy while using fewer training parameters, and has a stronger
anti-overfitting ability.
In this subsection, the recognition speed and the accuracy of the proposed algorithm are
compared with other recognition algorithms. All the experiments were based on the PLAID dataset,
with a Win10 system as the experimental environment, Python 3.8 as the language, and PyTorch
1.11.0 as the deep learning framework. 1074 sample data from 11 different types of household loads
are identified, and the ratio of the training set to the test set is still 8:2. Table 3 shows the
Even though only using V-I trajectories for recognition takes less time, the recognition
accuracy is poorer. Despite the K-means clustering method with a simple algorithm structure, the
identification accuracy isn't high, because no fusion features don't have a strong clustering effect.
Therefore, the recognition rate is still lower than the proposed algorithm. The step-by-step
recognition method in this paper not only effectively improves the recognition efficiency, but also
reduces the recognition time by combining the Inception network and TOPSIS.
To verify the noise resistance of the proposed algorithm, White Gaussian Noise is
superimposed on the raw voltage and current. The new voltage and current are obtained by adding
noise with different signal-to-noise ratios (SNRs), and the SNR is calculated by Equation (12).
L
(x ) i
2
SNR 10lg i 1
L
(12)
(ni )2
i 1
Where xi represents the original signal, ni represents the noise signal, and L is the length
Usually, the SNR indicates the deviation between the noise-combined signal and the original
signal, and the higher the SNR increases, the smaller the noise signal is.
The confusion matrixes of the first step by the Inception network under different noises with
SNRs of 5dB, 10dB, 20dB, and 30dB are shown in the Appendix. And the accuracy in the first step
keeps getting better as the SNR rises, as shown in Fig. 13. The identification rate of the Inception
network is still higher than the conventional CNN under different SNRs.
Fig. 13. Comparison of Recognition Accuracy for Conventional CNN and Inception under
Different SNRs
Under different SNRs, the step-by-step identification results are shown in Table 4. The step-
by-step identification algorithm can still achieve effective recognition of loads after adding noise,
and the recognition rate of the step-by-step model keeps improving as the SNR increases, which
indicates that the step-by-step model of the Inception-TOPSIS algorithm has a certain resistance to
noise.
Table 4. Recognition Accuracy under Different SNRs
SNR (dB) 5 10 20 30 40
Accuracy (%) 81.34 86.97 90.03 91.27 94.46
6. Conclusion
This paper proposes a step-by-step non-invasive loads identification method based on fused
features and Inception-TOPSIS. The step-by-step non-invasive loads identification method greatly
reduces computational difficulty, saves computational resources, and provides the possibility to
realize plug-and-play devices. The proposed method's effectiveness was fully validated through
experimental validation and comparative analysis using the public dataset PLAID dataset.
However, there are still some limitations to the proposed method. For instance, this paper only
has carried out the identification based on the public dataset, but the proposed method is not
validated based on the actual household datasets. In addition, an operational algorithm should be
tested in practical applications. Therefore, the next research focus is to solve the problems to
improve the scalability of the algorithm and try to use the embedded system to achieve real-time
References
[1] K. He, L. Stankovic, J. Liao, V. Stankovic, Non-intrusive load disaggregation using graph signal
processing, IEEE Transactions on Smart Grid. 9 (3) (2018) 1739-1747, doi:
10.1109/TSG.2016.2598872.
[2] Aladesanmi, E. J., K. A. Folly, Overview of non-intrusive load monitoring and identification
techniques, IFAC-PapersOnLine. 48 (30) (2015) 415-420, doi: 10.1016/j.ifacol.2015.12.414.
[3] Ahmed Zoha, Alexander Gluhak, Muhammad Ali Imran, Sutharshan Rajasegarar, Non-intrusive load
monitoring approaches for disaggregated energy sensing: A survey, Sensors. 12 (12) (2012) 16838-
16866, doi: 10.3390/s121216838.
[4] H. Y. Lam, G. S. K. F. , W. K. Lee , A Novel Method to Construct Taxonomy of Electrical Appliances
Based on Load Signatures, IEEE Transactions on Consumer Electronics. 53 (2007) 653-660,
doi: 10.1109/TCE.2007.381742.
[5] Z. Shi, B. Yin, An improved Non-intrusive load identification method for V-I trajectory based on
amplitude to pixel value, 2021 IEEE 5th Information Technology, Networking, Electronic and
Automation Control Conference (ITNEC). (2021), 1510-1516, doi:
10.1109/ITNEC52019.2021.9587263.
[6] Y. Liu, X. Wang, W. You, Non-Intrusive Load Monitoring by Voltage–Current Trajectory Enabled
Transfer Learning, IEEE Transactions on Smart Grid. 10 (5) (2019) 5609-5619, doi:
10.1109/TSG.2018.2888581.
[7] Du, L., He, D., Harley, R. G., Habetler, T. G, Electric Load Classification by Binary Voltage–Current
Trajectory Mapping. IEEE Transactions on Smart Grid. 7 (1) (2016) 358-365, doi:
10.1109/TSG.2015.2442225.
[8] J. Huan,C. Wang, H. Hong, Y. Sui, M. Yu, X. Pan, Non-intrusive load monitoring method based on
color encoding and deep learning, Science Technology and Engineering. 21(21)(2021) 8901-8908.(In
Chinese)
[9] C. Chen, P. Gao, J. Jiang, H. Wang, P. Li, S. Wan, A deep learning based non-intrusive household
load identification for smart grid in China, Computer Communications. 177 (2021) 176-184, doi:
10.1016/j.comcom.2021.06.023.
[10] X. Qiu, S. Yin, Z. Zhang, Z.Xie, M. Jiang, J. Zheng, Non-intrusive load identification method based
on V-I trajectory and high-order harmonic feature, Electric Power Engineering Technology. 46 (6)
(2021) 34-42, doi: 10.12158/j.2096-3203.2021.06.005. (In Chinese)
[11] W. Liu, C. Wang, Y. Li, Aggregation Method of Distributed Load Resources Based on Non-intrusive
Load Identification, Journal of Physics: Conference Series. 2138 (2021) 24-25, doi: 10.1088/1742-
6596/2138/1/012004.
[12] M. Yu, B. Wang, L. Lu, Z. Bao, D. Qi, Non-intrusive adaptive load identification based on Siamese
network, IEEE Access. 2022, 10 (2022) 11564-11573, doi: 10.1109/ACCESS.2022.3145982.
[13] S. Wang, L. Guo, H. Chen, X. Deng, Non-intrusive Load Identification Algorithm Based on Feature
Fusion and Deep Learning, Automation of Electric Power Systems. 44 (9) (2019) 103-110, doi:
10.7500/AEPS2019 0625010. (In Chinese)
[14] X. Xie, Z. Du, C. Hu, K. Yang, A. Wang, Reconfigurable design of Inception network, Computer
engineering and design. 43 (4) (2022) 1195-1201. (In Chinese)
[15] G. Qi, M. He, Convolutional Neural Network Image Classification Method Combined with Inception
Module, Software Guide. 19 (3) (2020) 79-82. (In Chinese)
[16] A. Wang, L. Yuan, C. Ding, G. Cao, P. Gao, Design and implementation of a non-intrusive power
load monitoring system, Journal of Xi’an University of Technology. 35 (3) (2019) 343-351, doi:
10.19322/j.cnk i.issn.1006-4710.2019.03.012. (In Chinese)
[17] J. Gao, E. C. Kara, S. Giri, M. Bergés, A feasibility study of automated plug-load identification from
high-frequency measurements, 2015 IEEE Global Conference on Signal and Information Processing
(GlobalSIP). (2015) 220-224, doi: 10.1109/GlobalSIP.2015.7418189.
[18] Lin M, Chen Q, Yan S, Network In Network, Computer Science. (2013).
[19] Q. Zhou, J. Wei, M. Sun, C. Wang, J. Rong, J. Hu, T. Yang, Feature Extraction for Non-intrusive Load
Monitoring System, 2021 6th Asia Conference on Power and Electrical Engineering (ACPEE). (2021)
503-507, doi: 10.1109/ACPEE51499.2021.9436971.
[20] C. Tian, Q. Zhang, G. Sun, Z. Song, S. Li, FFT consolidated sparse and collaborative representation
for image classification, Arabian Journal for Science and Engineering. 43 (2) (2017) 741-758, doi:
10.1007/s13369-017-2696-7.
[21] J. Li, H. Zhao, Power efficiency evaluation model for industrial users based on improved TOPSIS,
2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference
(ITAIC). (2020) 1764-1768, doi: 10.1109/ITAIC49862.2020.9338785.
[22] J. Huang, Combining entropy weight and TOPSIS method for information system selection, 2008
IEEE conference on cybernetics and intelligent systems. (2008) 1281-1284, doi:
10.1109/ICCIS.2008.4670971.
[23] J. Yin, X. Du, H. Yuan, M. Ji, X. Yang, S. Tian, Q. Wang, Y. Liang, TOPSIS Power Quality
Comprehensive Assessment Based on A Combination Weighting Method, 2021 IEEE 5th Conference
on Energy Internet and Energy System Integration (EI2). (2021) 1303-1307, doi:
10.1109/EI252483.2021.9713201.
[24] R. Medico, L. D. Baets, J. Gao, S. Giri, E. Kara, T. Dhaene, C. Develder, M. Bergés, D. Deschrijver.
A voltage and current measurement dataset for plug load appliance identification in
households, Scientific data. 7 (1) (2020) 1-10, doi: 10.1038/s41597-020-0389-7.
[25] De Baets, Leen,J. Ruyssinck,C. Develder,T. Dhaene,D. Deschrijver, Appliance classification
using VI trajectories and convolutional neural networks, Energy and Buildings. 158 (2018) 32-36,
doi: 10.1016/j.enbuild.2017.09.087.
[26] F. Jazizadeh,B. Becerik-Gerber,M. Berges,L. Soibelman, An unsupervised hierarchical
clustering based heuristic algorithm for facilitated training of electricity consumption disaggregation
systems, Advanced Engineering Informatics. 28 (4) (2014) 311-326, doi: 10.1016/j.aei.2014.09.004.
Appendices