See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/272279159

Load Identification in Neural Networks for a Non-

intrusive Monitoring of Industrial Electrical Loads

Article  in  Lecture Notes in Computer Science · January 2008

CITATIONS READS
36 284

3 authors:

Hsueh-Hsien Chang Hong-Tzer Yang

JIN WEN UNIVERSITY OF SCIENCE AND TECHNOLOGY National Cheng Kung University
46 PUBLICATIONS   788 CITATIONS    59 PUBLICATIONS   2,235 CITATIONS   

SEE PROFILE SEE PROFILE

Ching-Lung Lin
Minghsin University of Science and Technology
48 PUBLICATIONS   1,127 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Study of Constructions for the Photovoltaic System to Increase the Economic Efficiency of Energy Generation View
project

All content following this page was uploaded by Hsueh-Hsien Chang on 15 February 2015.

The user has requested enhancement of the downloaded file.

 Load Identification in Neural Networks for a
Non-intrusive Monitoring of Industrial Electrical Loads

Hsueh-Hsien Chang1,2, Hong-Tzer Yang2, and Ching-Lung Lin3

1
Department of Electronic Engineering, Jin Wen University of Science and Technology, Taipei,
Taiwan
sschang@just.edu.tw
2
Department of Electrical Engineering, Chung Yuan Christian University,
Taoyuan, Taiwan
htyang@dec.ee.cycu.edu.tw
3
Department of Electrical Engineering, Ming Hsin University of Science and Technology,
Hsinchu, Taiwan
cll@must.edu.tw

Abstract. This paper proposes the use of neural network classifiers to evaluate
back propagation (BP) and learning vector quantization (LVQ) for feature se-
lection of load identification in a non-intrusive load monitoring (NILM) system.
To test the performance of the proposed approach, data sets for electrical loads
were analyzed and established using a computer supported program - Electro-
magnetic Transient Program (EMTP) and onsite load measurement. Load iden-
tification techniques were applied in neural networks. The efficiency of load
identification and computational requirements was analyzed and compared using
BP or LVQ classifiers method. This paper revealed some contributions below.
The turn-on transient energy signatures can improve the efficiency of load iden-
tification and computational time under multiple operations. The turn-on tran-
sient energy has repeatability when used as a power signature to recognize
industrial loads in a NILM system. Moreover, the BP classifier is better than the
LVQ classifier in the efficiency of load identification and computational
requirements.

Keywords: Load identification, neural network, non-intrusive load monitoring,

Electromagnetic Transient Program.

1 Introduction
Traditional load-monitoring instrumentation systems employ meters for each load to be
monitored. These meters may incur significant time and cost to install and maintain.
Furthermore, increasing numbers of meters may impact system reliability. Therefore, a
method for minimizing the number of instruments using non-intrusive load monitoring
(NILM) is needed.
To develop such a monitoring system, a number of load identification techniques
have been proposed [1-10]. Hart proposed several load identification methods that
examined the steady-state behavior of loads [1]. Hart conceptualized a finite state

W. Shen et al. (Eds.): CSCWD 2007, LNCS 5236, pp. 664–674, 2008.
© Springer-Verlag Berlin Heidelberg 2008
 Load Identification in Neural Networks for a Non-intrusive Monitoring 665

machine to represent a single appliance in which power consumption varied discretely

with each step change. Leeb used finite-impulse response filters [3-6] and Robertson
employed a wavelet transformation technique [7] matched to transient envelopes in
particular load-switching models. Cole [8] examined data extraction methods and
steady-state load identification algorithms for NILM. The algorithms developed by
Cole could be employed for load switching between individual appliances when one or
more appliances are switched on or off.
However, some appliances have transient features and steady-state features with
quite different indications and meaning at different times, particularly in an industrial
plant. For example, appliances may have variable power during operation, such as
water pumps in a steel mill. These appliances have no sufficient features to identify
load which is starting or operating, if these features are only steady-state features or
transient features. These features, when selected, enhance the capability of a NILM
system to distinguish from industrial loads under multiple operations using the turn-on
transient energy feature (UT) and traditional steady-state power features by analyzing
the physical industrial load characteristics. They also improve computational time.
In an industrial environment, an Electromagnetic Transient Program (EMTP)
computer supported program allows convenient comparison of load prototypes or
templates under normal load operation. To maximize recognition accuracy, analysis of
turn-on transient energy signatures uses a window of samples, Δt, to differentiate
adaptively a transient representative class of loads. The experiments in this study reveal
some contributions below. The turn-on transient energy signatures can improve the
efficiency of load identification and computational time under multiple operations. The
turn-on transient energy has repeatability when used as a power signature to recognize
industrial loads in a NILM system. Moreover, the back propagation (BP) classifier is
far better than the learning vector quantization (LVQ) classifier, within the efficiency
of load identification and computational requirements.
This paper is organized as follows. The power signature problems and turn-on
transient energy algorithm are addressed and described in Section 2. The event detec-
tion is described in Section 3. The turn-on transient energy repeatability is described in
Section 4. Based on the algorithm and load recognition techniques, a series of load
recognition experiments for the feature selection of power signals are conducted in
Section 5, which also includes comparisons of recognition accuracy and computational
performance using selected features by turn-on transient energy algorithm and tradi-
tional power signal features under individual operation or multiple operations. Ad-
vantages of load recognition using the turn-on transient energy algorithm and BP are
concluded in Section 6.

2 Power Signature Problems and Turn-On Transient Energy

Algorithm

2.1 Power Signature Problems for Feature Selection

In general, an appliance may have many load representations and a load may involve
many physical components. For example, a dryer has two loads, a motor and a heater. A
refrigerator has only one load, a compressor, but has different physical components
for defrosting and freezing. Most appliances are distinguishable by unique power
666 H.-H. Chang, H.-T. Yang, and C.-L. Lin

signatures that can be observed from voltage and current waveforms supplied to the
appliance, or from processed reproductions of these signals such as the delivered real
power (P) and reactive power (Q) or harmonics [2].
According to the switch continuity principle, steady-state signatures, for example,
real power and reactive power, are additive when two signatures occur simultaneously.
In contrast to steady-state properties, transient properties are not addition [1]. Distin-
guishing different loads may be problematic when they have equivalent real power and
reactive power but no harmonic components, and/or when the sums of real power and
reactive power of two load types are equal to that of another load during multiple load
operations. Since most steady-state data is processed in batch format within a period of
one or a few days, they cannot make real-time identification for load operation. Load
detection, based on steady-state power, makes an industrial NILM system, which is
susceptible to confusion, if two loads start up at nearly the same time.

2.2 Turn-On Transient Energy Algorithm

The transient properties of a typical electrical load are mainly determined by the
physical task that the load performs [4-6]. Transient energy may assume different forms
in consumer appliances, depending on the generating mechanism [1]. Estimating
current waveform envelopes at the utility service entry of a building, for example,
allows accurate transient event detection in the NILM [5]. Load classes performing
physically different tasks are therefore distinguishable by their transient behavior [4-6].
Since the envelopes of turn-on transient instantaneous power are closely linked to
unique physical quantities, they can serve as reliable metrics for load identification.
Two different appliances consuming identical levels of real power and reactive power
may have very different turn-on transient currents. Analysis of these transient currents
can accurately determine which of the two is actually present in the load.
In general, the transient behavior of many important loads is sufficiently distinct
reliably to identify load type. The long characteristic switching-on transient, the less
substantial switching-on transient, the short but very high-amplitude switching-on
transient and the long two-step switching-on transient are the principal values measured
in pump-operated appliances, motor-driven appliances, electronically fed appliances
and fluorescent lighting, respectively [11].
During training phases, a window of samples of time length Δt is examined to dif-
ferentiate transients representative of a class of loads. This segmentation process
delineates a set of transient energy values representing a particular transient shape in
∧
each of the input envelopes. To maximize the recognition accuracy ( γ ) during the test
phase, Δt is adaptively changed based on factor δ. Figure 1 shows the algorithm adapted
for Δt by factor δ. As demonstrated in Eq. 4, the search for a precise time pattern for
instantaneous power turn-on identifies a complete transient.

Q dWTransient = v ⋅ dq . (1)

dq = i ⋅ dt . (2)

∴ dWTransient = v ⋅ i ⋅ dt . (3)
 Load Identification in Neural Networks for a Non-intrusive Monitoring 667

t s + Δt
WTransient = ∫ v ⋅ i ⋅ dt . (4)
ts

dWTransient (t )
PIns tan tan eous (t ) = = v(t ) × i (t ) . (5)
dt

v, i, Δ t, γ
∧
initial

Turn-on Transient Event Detector for ts

ts + Δ t
W Transient = ∫t s v ⋅ i ⋅ dt

Energy Measurement
Δ t = Δt + δ
Neural Network Algorithms
δ < Δt

No
∧
γ > γ

Yes

∧
γ ⇐ γ

Fig. 1. Adaptive algorithm for Δt by factor δ at the maximum recognition accuracy of the turn-on
transient energy

3 Event Detection
Figure 2 shows the measured envelope of instantaneous power in one phase of the
turn-on transient of a load bank connected to an A.C. source via a six-pulse power
electronic converter. Completion of the turn-on transient in this load bank is followed
by the turn-on transient of a three-phase 140-hp induction motor. This study proposes
the use of wavelet transformation analysis to detect and localize various turn-on tran-
sient events. The underlying concept of this approach is decomposition of a given
turn-on transient signal into other signals to enhance the original signal, including
shape edges, transitions and jumps [7, 12]. The approach also determines whether the
wavelet transformation coefficients (WTC) are unique, since they represent the
occurrence of the turn-on event.
Figure 3 illustrates event detection using the Daub2 wavelet transformation at scale
1 for the envelope of the instantaneous turn-on power signal in Fig. 2. Figure 3 shows a
detailed illustration of sharp turn-on transient events occurring in the vicinity of 0.05
and 0.15 s. The WTC at scale 1 show the occurrence of the sharp events of a turn-on
transient. Maximum localization of the analytical wavelet is at scale 1.
668 H.-H. Chang, H.-T. Yang, and C.-L. Lin

Fig. 2. The envelope of overlapping instantaneous power in one phase of the turn-on transient of
a load bank connected to a six-pulse power electronic converter. Completion of the turn-on
transient in a load bank is followed by the turn-on transient of a three-phase 140-hp induction
motor.

Fig. 3. Event detection uses the Daub2 wavelet transformation of the detailed version at scale1
for the envelope of the instantaneous turn-on power signal in Fig. 2

4 Turn-On Transient Energy Repeatability

Most loads observed in the field have repeatable transient profiles or at least repeatable
sections of transient profiles [3]. The load survey reveals that non-linearity in the consti-
tutive relationships of the elements comprising a load model and/or in the state equation
describing a load tends to produce interesting and repeatable observable turn-on transient
profiles suitable for use in identifying specific load classes [4-6]. Because of the varying
transients (which often depend on the exact point in the voltage cycle at which the switch
opens or closes), data sets for load identification must provide accurate repeatability of the
turn-on transient energy signatures. Pre-training has proven effective for highly repeatable
loads that occur frequently, such as commercial and industrial loads.
As Figures 4 and 5 demonstrate, the turn-on transient profiles exhibit repeatable
measured current waveforms in one phase at voltage phase 00 and 900 for the turn-on
transient of a three-phase 300-hp induction motor. The turn-on characteristics of a load
clearly increase in complexity over time. Closer investigation of the load turn-on is thus
required before the characteristics can be used as a distinguishing feature of a load. This
information, collected via non-intrusive monitoring, can be used to answer important
questions about the statistical validity of power measurements.
Determination of whether or not turn-on transient energy content is repeatable
would be useful in developing a turn-on transient energy signature. As Eq. (6) shows,
the average value of the sample data ( x i ) for the turn-on transient energy of each load
is x . The standard deviation (S) of the turn-on transient energy for each load is com-
puted according to Eq. (7) for all loads monitored in isolation. An experiment is then
used to demonstrate that the statistical validity of the turn-on transient energy for each
load is repeatable in terms of the coefficient of variation (C.V.) according to Eq. (8).
 Load Identification in Neural Networks for a Non-intrusive Monitoring 669

Fig. 4. Current waveform in one phase at voltage phase 00 for the turn-on transient of a
three-phase 300-hp induction motor

Fig. 5. Current waveform in one phase at voltage phase 900 for the turn-on transient of a
three-phase 300-hp induction motor

1 n
x = ∑ xi . (6)
n i =1

1
∑ (x i − x )2
n
s = . (7)
n − 1 i =1
s (8)
C .V . = .
x

5 Experimental Results and Discussions

5.1 Experimental Environment

The NILM system in this paper monitors the voltage and current waveforms in a
three-phase electrical service entry powering representative loads of important load
classes in an industrial building. The neural network algorithm in the NILM system
identifies five actual loads with transient and steady-state signatures at the 480-V
common bus. These loads include a 95-hp induction motor, a 300-hp synchronous
motor, a 140-hp induction motor, a 95-hp induction motor driven by line frequency
variable-voltage drives and a bank of loads supplied by a six-pulse power electronic
converter for A.C. power.
Figure 6 schematically illustrates the test stand used in experiment. Three-phase
480-V electricity powers the loads, which are representative of important load classes
in an industrial building. A dedicated computer connected to the circuit breaker panel
controls the operation of each load. The computer can also be programmed to stimulate
various end-use scenarios. To compile data for training purposes, either every site of
interest or a representative sample of the sites should be monitored. The sample rate is
approximately 15 kHz. The appropriate coefficients corresponding to the current and
670 H.-H. Chang, H.-T. Yang, and C.-L. Lin

voltage in each harmonic are extracted from the results. For this study, a neural network
simulation program was designed using MATLAB. This program was executed on an
IBM PC with an Intel 1.5-GHz Pentium M CPU for load identification.

Host PC
Substation

22.8kV/ 480V
3MVA, 6% Dyn
Non-intrusive
Load Monitoring 3φVoltage Measurements
System
3φ Current Measurements

480V Bus

PC
1 2 3

4 3φ 6 pulses
Rectifer
3φ95hp,4poles 3φ300hp 3φ140hp,4poles Rload =4Ω
squirrel-cage I/M 2poles,S/M squirrel-cage I/M
3φ95hp,4poles
squirrel-cage I/M
Line Frequency
Variable-Voltage
Drives

Fig. 6. Electrical schematic of NILM in an industrial plant

5.2 Turn-On Transient Energy Repeatability

To determine whether turn-on transient energy exhibits repeatability, a NILM

system with five important loads was examined in an industrial plant, as shown in
Figure 6. The turn-on transient energy for each load can be computed from the
measured voltage and current waveform at the service entry according to Eq. (4).
Because of the varying transients (which often depend on the exact point in the
voltage cycle at which the switch opens or closes), it is essential that data sets for
load identification have highly repeatable transient energy signatures. Therefore, the
instantaneous power profile for each turn-on transient load is sampled when the
system utility voltage is switched from 00 to 3500 at 100 intervals, i.e., the number of
load samples (n) is 36.
Table 1 shows C.V. values during periods of nearly steady energy for each load,
all of which are less than 1% [4]. The simulation results indicate that the turn-on
transient energy should have good repeatability. Therefore, the turn-on transient
energy can be used as a power signature to recognize industrial loads in a NILM
system.
 Load Identification in Neural Networks for a Non-intrusive Monitoring 671

Table 1. Variation coefficient during periods of nearly steady energy for each load in a NILM
system

Loads Load 1 Load 2 Load 3 Load 4 Load 5

C.V. (%) 0.36648 0.02245 0.16077 0.47399 0.10885

5.3 Load Identification in Individual Operation

Changing circuit breaker number 1 to number 5 in individual operation yields data for
individual load. Five combinations are possible. Each data set includes a voltage
variation from –5% to +5% at 1% intervals. The total number of data sets is 55 (5×11).
To confirm the inferential power of the neural network, the data are categorized as 28
training and 27 test data sets. Notably, training data and test data are selected randomly
from all data.
Table 2 shows that values for the training accuracy of load identification in indi-
vidual operation are 100% for features with real power and reactive power, and/or with
total harmonic distortion and/or with turn-on transient energy for one of the features.
Furthermore, the test accuracy of load identification in individual operation is at least
96%.
Except when the features are real power and reactive power, table 3 shows that
values for the training and test recognition accuracy of load identification in individual
operation are 100% for features with real power and reactive power, and with total
harmonic distortion and/or with turn-on transient energy for one of the features.

Table 2. Recognition accuracy and computational requirements of back propagation classifier

1 2
Features PQ PQVTHD ITHD PQVTHDITHDUT
Training Testing Training Testing Training Testing
Recognition 100 96.43 100 100 100 100
Accuracy (%)
Computational 5.348 1.752 4.006 1.773 2.219 1.297
Time (s)
The Number 7 9 10
of Hidden
Neurons

5.4 Load Identification in Multiple Operations

Changing circuit breaker number 1 to number 5 in multiple operations yields data for
multiple loads. Thirty-one combinations are possible. Each data set includes a voltage
variation from –5% to +5% at 1% intervals. The total number of data sets is 341
(31×11). To confirm the inferential power of the neural network, the data are catego-
rized as 171 training and 170 test data sets. Notably, training data and test data are
selected randomly from all data.

1
VTHD is the total voltage harmonic distortion.
2
ITHD is the total current harmonic distortion.
672 H.-H. Chang, H.-T. Yang, and C.-L. Lin

Table 3. Recognition accuracy and computational requirements of learning vector quantization

classifier

Features PQ PQVTHDITHD PQVTHDITHDUT

Training Testing Training Testing Training Testing
Recognition 81.48 78.57 100 100 100 100
Accuracy (%)
Computational 15595 0.266 59.526 0.250 12.939 0.24
Time (s)
The Number 1000 250 25
of Hidden
Neurons

Table 4 shows that values for the training accuracy of load identification in multiple
operations are 100% for features with real power and reactive power, and/or with total
harmonic distortion and/or with turn-on transient energy for one of the features. Fur-
thermore, the test accuracy of load identification in multiple operations is at least 94%.
Table 5 shows that values for the training accuracy of load identification in multiple
operations are 42.1% for features with real power and reactive power, and with total
voltage/current harmonic distortion and with turn-on transient energy for one of the
features. Furthermore, the test accuracy of load identification in multiple operations
with the same features is 42.94%, which is higher than other features.

Table 4. Recognition accuracy and computational requirements of back propagation classifier

Features PQ PQVTHDITHD PQVTHDITHDUT

Training Testing Training Testing Training Testing
Recognition 100 94.73 100 94.73 100 95.3
Accuracy (%)
Computational 313.471 1.843 293.640 1.893 246.719 1.516
Time (s)
The Number 7 9 23
of Hidden
Neurons

5.5 Discussions

There are different representations of load patterns for individual operation and mul-
tiple operations. In individual operation, a class shows that the representation is only
one load. In multiple operations, a class shows that the representation can be one or
many loads. In other words, a class may be a combination of more than one load.
Therefore, classifications are more complicated for multiple operations.
These features cannot be adequately measured only from steady-state parameters in
multiple operations, that is, real power and reactive power. In other words, it is difficult
in steady-state power to identify each load, when the sums of real power and reactive
power of two loads types are equal to that of another load. In contrast to steady-state
properties, transient properties such as the turn-on transient energy can play an
 Load Identification in Neural Networks for a Non-intrusive Monitoring 673

Table 5. Recognition accuracy and computational requirements of learning vector quantization

classifier

Features PQ PQVTHDITHD PQVTHDITHDUT

Training Testing Training Testing Training Testing
Recognition 2.34 2.94 20.5 22.4 42.10 42.94
Accuracy (%)
Computational 55017 0.765 56772 0.938 59477 0.969
Time (s)
The Number 15000 15000 15000
of Hidden
Neurons

important role. Combining transient and steady-state signatures is necessary to improve

recognition accuracy for NILM.
In the individual operation, back propagation classifier requires less computational
requirements than learning vector quantization classifier. In multiple operations, the
computational requirements can not compare with the back propagation classifier and
the learning vector quantization classifier, for the algorithm of LVQ classifier can not
converge until the number of maximum iterations.

6 Conclusions
The EMTP simulation is invaluable for testing pattern recognition samples and allows
the rapid development and implementation of successful prototypes. The NILM system
employs an adaptive algorithm of the turn-on transient energy for start-up analysis to
improve the efficiency of load identification and computational time. The testing
recognition accuracy can be relatively high at 95.3% for back propagation classifier, in
multiple operations. Furthermore, the BP classifier is far better than the LVQ classifier,
in the efficiency of load identification and computational requirements.

References
1. Hart, G.W.: Non-intrusive Appliance Load Monitoring. In: Proc. 1992 IEEE Conference,
vol. 80, pp. 1870–1891 (1992)
2. Roos, J.G., Lane, I.E., Botha, E.C., Hanche, G.P.: Using Neural Networks for Non-intrusive
Monitoring of Industrial Electrical Loads. In: Proc. 1994 IEEE Instrumentation and Meas-
urement Technology Conference, pp. 1115–1118 (1994)
3. Laughman, C., Lee, K., Cox, R., Shaw, S., Leeb, S.B., Norford, L., Armstrong, P.: Power
Signature Analysis. IEEE Power & Energy Magazine, 56–63 (2003)
4. Norford, L.K., Leeb, S.B.: Non-intrusive Electrical Load Monitoring in Commercial
Buildings based on Steady-state and Transient Load-detection Algorithm. Energy and
Buildings 24, 51–64 (1996)
5. Leeb, S.B., Shaw, S.R., Kirtly Jr., J.L.: Transient Event Detection in Spectral Envelop es-
timates for Nonintrusive Load Monitoring. IEEE Trans. Power Delivery 10(3), 1200–1210
(1995)
674 H.-H. Chang, H.-T. Yang, and C.-L. Lin

6. Leeb, S.B.: A Conjoint Pattern Recognition Approach to Non-intrusive Load Monitoring.

Ph. D. Dissertation. Department of Electrical Engineering and Computer Science, M. I. T.
(1993)
7. Robertson, D.C., Camps, O.I., Mayer, J.S., Gish, W.B.: Wavelets and Electromagnetic
Power System Transients. IEEE Trans. Power Delivery 11(2), 1050–1057 (1996)
8. Cole, A.I., Albicki, A.: Nonintrusive Identification of Electrical Loads in a Three-phase
Environment based on Harmonic Content. In: Proc. IEEE Instrumentation and Measure-
ment Technology Conference, pp. 24–29 (2000)
9. Yang, H.-T., Chang, H.-H., Lin, C.-L.: Design a Neural Network for Features Selection in
Non-intrusive Monitoring of Industrial Electrical Loads. In: Proceedings of the 11th Inter-
national Conference on Computer Supported Cooperative Work in Design, vol. 2, pp.
1022–1027 (2007)
10. Murata, H., Onoda, T.: Applying Kernel Based Subspace Classification to a Non-intrusive
Monitoring for Household Electric Appliances. In: Dorffner, G., Bischof, H., Hornik, K.
(eds.) ICANN 2001. LNCS, vol. 2130. Springer, Heidelberg (2001)
11. Sultanem, F.: Using Appliance Signatures for Monitoring Residential Loads at Meter Panel
Level. IEEE Trans. Power Delivery 6(4), 1380–1385 (1991)
12. Santoso, S., Powers, E.J., Grady, W.M., Hofmann, P.: Power Quality Assessment via
Wavelet Transform Analysis. IEEE Trans. Power Delivery 11(2), 924–930 (1996)

View publication stats