Tutorial:ANN:sin power systems

Artificial neural networks

in power systems
Part 3 Examples of applications
in power systems
This tutorial describes some typical applications of artificial neural networks
(ANNs) in power systems. It is the third in a series of three articles which, through
a consideration of real problems, illustrates some of the practical aspects of ANN
design in terms of architecture, training data requirements, selection of input
features and learning algorithms.

by Raj Aggarwal and Yonghua Song

mainstream of applications in power systems;

P
arts 1 and 2 of this series of three
articles outlined, respectively, the this is principally because the supervised
underlying concepts of ANNs and a learning associated with the MLP is superior in
description of the different types of terms of accuracy compared with either the KN
ANNs in use today. ANNs represent a growing or HN. However, as shown later, it is important
new technology as indicated by the wide variety to appreciate that unsupervised networks, such
of applications, such as remote sensing, as the KN, are being increasingly employed as a
control, forecasting, pattern recognition etc. data set preprocessing step; this reduces thLe size
The main reasons for this growing activity are of training data requirements before the
the ability of ANNs to learn complex non-linear network handles the data characteristics with
relationships and their modular structures, supervised learning. The overall training of the
which allow parallel processing. ANNs have ANN is also significantly expedited.
been shown to be useful in solving algorithmic- There are now widespread applications of 1 Characteristics of
type problems and, more importantly, in ANNs in power systems. However, this tutorial power system problems
dealing with problems for which algorithms are
not available but significant data is at hand.
In the areas of power systems, problems and
their formulation may be expressed in different
ways depending on their nature, as shown in
Fig.1. Among these factors, the non-linear and
large-scale systems generally make power
system problems more difficult to solve. In this
respect, no suitable analytical technology is
available to deal with this complexity and an
ANN offers a promising and attractive tool for
dealing with this level of intricacy
Of the three most common types of ANNs,
namely multi-layer, perceptron (MLP),
Kohonen network (KN) and Hopfield network
(HN), the MLP has hitherto been the

POWER ENGINEERING JOURNAL DECEMBER 1998 279

Tutorial: ANNs in power systems

deals with only a few specific problems namely, approaches. One treats the load demand as a
load forecasting, fault classification, voltage time series signal and predicts the load using
stability, economic dispatch and design of different time-series analysis techniques. The
voltage stabilisers, in some detail. Emphasis is second method is a regression technique which
placed on some practical aspects of ANN recognises the fact that the load demand is
design, in particular: the selection of the heavily dependent on weather conditions.
appropriate ANN architecture; the selection of However, such traditional techniques often do
a training data set in terms of both training not give sufficientlyaccurate results. One major
requirements and the selection of the most drawback of the time series is that it does not
significant features; the learning scheme; and utilise weather conditions. The main problem
finally performance evaluation. The goal is with the regression technique is that it
accomplished by presenting and analysing data unjustifiably assumes a linear relationship
attained from typical power systems. between weather conditions and load demand.
Conversely,complex algorithmic methods with
Short-term load forecasting using a multi- heavy computational burden can slow the
layer perceptron convergence and in certain cases may cause
Forecasting electrical loads in a power system divergence.
with lead times varying from a few minutes to One very attractive approach to the problem
hours to days has obvious economic as well as of accurate load forecast is using an ANN; this
other advantages. The forecast information can is because an ANN can combine both time
be used to aid optimal energy interchange series and regression approaches to predict the
between utilities, thereby saving valuable fuel load demand. A functional relationship
costs. Forecasts also significantly influence between weather variables and electrical load is
important system operations decisions such as not needed. This is because the ANN can
dispatch, unit commitment and maintenance technically generate this functional
scheduling. For these reasons, considerable relationship by learning and training data. In
efforts are being invested in the development of other words, the non-linear mapping between
accurate load forecasting techniques. the inputs and outputs is implicitly embedded
In this article, we will concentrate on the in the ANN.
problem of short-term forecasting up to 24 We will now discuss one specific ANN-based
hours ahead. These load forecasts play a crucial load forecasting approach which combines
role in optimum unit commitment, start up and both time series and regression approaches.
shut down of thermal plant, control of spinning The algorithm utilises an ANN based on the
reserve, and buying and selling of power in MLP architecture. As is the case with the time-
of a three- interconnected systems. series approach, the ANN traces previous load
layered perceptron type Most of the conventional techniques used in patterns and predicts (i.e. extrapolates) a load
ANN load forecasting can be categorised under two pattern using recent load data. The algorithm
also uses weather information for modelling
and the ANN is able to perform non-linear
modelling and adaptation. One of the principle
advantages over conventional techniques is
that it does not require the assumption of any
functional relationship between load and
weather variables in advance. The ANN can be
adapted by exposing it to new data.
As discussed in the second article (February
1998 PEJ), an ANN can be defined as a highly
connected array of elementary processes called
neurons. Fig. 2 is an example of the widely used
MLP-type ANN employed in the application
considered here. A total of Q sets of training
data are assumed to be available. Inputs .of (il,
iz, ..., i ~ are
) imposed on top (Le. input) layer.
The ANN is trained to respond to the
corresponding target vector (tl, t2, ..., te) on the
bottom (i.e. output) layer. The training

280 POWER ENGINEERING JOURNAL DECEMBER 1998

 Tutorial: ANNs in power systems

3 Load curves for

England and Wales

continues until a certain error criterion is simply one large ANN. The selected input/
satisfied. Typically, the training is halted when output variables for the ANN are given in Table
the average error between the desired and 1.
actual outputs of the ANN over the Q training The chosen input variables are very
data sets is less than a predetermined threshold. informative by virtue of the fact that they
The training time is dictated by various implicitly reflect possible weather changes. The
elements including the complexity of the number of training patterns presented to the
problem, the volume data, the network ANNs is important as it can have a signiificant
structure and the training parameters used. bearing on forecast accuracy. In this case study,
In the application considered here, an ANN 360 training patterns were used for the
is trained to recognise and predict the load weekday model and 290 for the weekend
demand for England and Wales. In this respect, model.
Fig. 3 typifies the load demand over a 24 hour The ANNs were comprised of the MLP and
period for four seasons. It is apparent that the were based on the standard error-back-
most obvious difference in load demand occurs propagation (EBP) training algorithm; the
in summer and winter. The daily peak load in epochs and the hidden neurons were varied
winter can be as high as 45 000 MW compared from 5000-15000 and 2-9, respectively The
to 33000 MW in summer. The differences in best MLP was obtained by comparative cross-
load demand for autumn and spring are not that validation amongst several MLPs. The
discernible since these seasons represent measurement of forecasting performances was
transitional periods in which there is a gradual
change in load conditions from one season to Table 1 Inputloutput variables for the ANN (24-
the other. Moreover, the weekday (Monday to hour lead time)
Friday) electricity demands are much higher
than those during weekends (Saturday and
Sunday).
Based on the analyses of load demand
patterns, a year is partitioned into four seasonal
periods with equal time intervals. Each
seasonal period is further divided into two day
types, weekdays and weekends. Independent where LM(d,t)= (L(d-7,t) + L(d-14,t) + L(d-21,t))./3
forecast models are then developed for each reflects the load demand on day d at hour t
season and for each day type by utilising only indicates the daily load changes
the load information of the respective group. provides the weekly trend of load demand
With this approach, the performance attained is changes on day d at hour t
much superior than would be the case if the makes use of the average of the same three
whole problem were to be encapsulated in days at hour t i n the previous 3 weeks

POWER ENGINEERING JOURNAL DECEMBER 1998 28 1

Tutorial: ANNs in ower systems

4 Forecast and actual Fault classification in double-circuit trans-

load profiles mission lines using combined unsupervised/
supervised ANNs
Parallel transmission lines which can
significantly increase transmission capacity on
existing systems are finding more widespread
usage. However, there is difficulty in classifying
the fault types on such lines using conventional
techniques, principally because faulted
phase(s) on one circuit have an effect on the
phases of the healthy circuit due to mutual
coupling between the two circuits.The problem
is compounded by the fact that this coupling is
highly non-linear in nature and is dependent on
a complex interplay amongst a number of
variables. As a consequence, the coupled
phase(s) on the healthy circuit may sometimes
be wrongly diagnosed as being faulted phase(s)
under certain fault conditions. Thus
conventional classifiers based on logical
comparison techniques or linear algorithms are
not well suited for such circuits. In this respect,
neural computing has the very important
attribute of being able to solve non-linear
system identification problems through using
neurons, links and learning algorithms, and
hence ANNs are ideally suited to deal with
complex non-linear fault classification
problem.
ANNs have to be trained to learn and, in this
respect, the training algorithms can be divided
into supervised, unsupervised and combined
unsupervisedlsupervised as shown in Fig. 5.
Classifiers trained with supervision require
data labels that specify the correct class during
training. Clustering algorithms use
unsupervised training and group unlabelled
training data into internal clusters. Classifiers
that use combined unsupervisedlsupervised
training firstly use unsupervised training with
unlabelled data to form internal clusters; labels
are then assigned to clusters during the
based on mean absolute percentage error supervision stage. Different ANNs with ,
(MAPE), where: different training techniques have their own
advantages and disadvantages. A typical
9
MAPE = 1/NC(Lfovecasti x 100
- Laciuali)/Lactuail supervised error back-propagation (EBP)
,=I
network is a non-linear regression technique
The trained models were then used to forecast which attempts to minimise the global error. An
the load demand over a 24-hour period. Fig. 4 EBP network can provide very compact
typifies the load demand, actual and forecast distributed representations of complex data
(with lead times of 24 hours), as predicted by sets, and is smaller in size compared with a
the ANNs, for two separate days in a week combined unsupervisedlsupervised ANN with
during the winter months. As can be seen, the the same inputs and outputs. However, training
ANN forecasts closely match the actual of an EBP network is very .slow (time
demand, the average MAPE for a weekday and consuming), needs much larger training sets
weekend being 1.2% and 1.5%,respectively. and it very easily gets stuck on local minima.

282 POWER ENGINEERING JOURNAL DECEMBER 1998

 Tutorial:ANN:s in power systems

Furthermore, it can be difficult to retrain the

ANN with new training data.
Unsupervised learning means learning
training
examples without teaching, i.e. there are no
desired outputs. A typical unsupervised
learning network is the KN. The network
attempts to learn a topological map from an N-
dimensional input space into a two
supervised combined unsupervised
dimensional feature space. Thus the network training unsupervised/
has many advantages over the EBP network, supervised
training
such as fast learning, a much smaller amount of
training data etc. But, in view of the fact the
network is without an output layer, it is not
recommended to be used on its own for either
pattern classification or other decision-making
processes. Rather, it is used as the front end to
an output layer with supervised learning and
end SI circuit 1 end S2
becomes a combined unsupervisedhpervised
(CUS) learning network, the subject of the
technique described here.
5 GVA +line length 128 i m -+ 35 GVA
This article proposes a novel fault type
classification technique for double-circuit
transmission lines using a CUS network. The
system studied is shown in Fig. 6 and is a
typical 400 kV double-circuit line of the type should be unsupervised in a sense, because the 5 ANN training
encountered in the UK. target class for each object is unknown. Fig. 7 techniques
The CUS-based classifier is a technique that typifies a CUS-based network.
6 The transmission
separates object recognition into two parts: (i) The input vector for the CUS network
systemstudied
feature extraction with unsupervised learning comprises nine variables associated with the
in the first stage, and (ii) classification with three voltage and six current signals in a
supervised learning sitting on the top, double-circuit line. The feature extraction is
subsequently. An important basic principle is based on time-domain windows, each window-
that the features must be independent of class length being of three samples. The outputs are
membership, since the latter is not yet known composedofsevenvariablesAl,Bl, C1,PL2,B2,
at the feature extraction stage by definition. C2 and G; of these, ‘1’and ‘2’ signify circuits 1
This implies that, if any learning methods are and 2, respectively, and variable G indlicates
used for developing the feature extractors, they whether ground is involved in a fault. As an

7 CUS-based network
architecture

POWER ENGINEERING JOURNAL DECEMBER 1998 283

Tutorial: ANNs in power systems

converged after only 4000 iterations, i e in

approximately 2 minutes, and reached a much
lower RMS error of 0 03 Furthermore
(although not shown here), of the 100 test cases
considered, the misclassification rate was 6%
and 1% for the MLP and CUS networks,
respectively Fig 8 typifies the RMS
convergence diagrams for the two ANNs, re-
inforcing the far superior performance of the
CUS network both in terms of accuracy and
training speed.

Kohonen neural network for voltage

stability analysis
Voltage stability is one of the major concerns in
8 RMS error diagrams example, if we get an output 1,0,0,0,0,0,1 this the operation and control of modern power
for the ANNs would indicate an ‘a’-phase-earth fault on systems. Very often, the voltage instability
considered
circuit 1. phenomenon is considered to be
In order to ascertain the attributes of the CUS geographically local. Some areas or buses are
network over a MLP network (which has an much more sensitive than others. A system may
architecture similar to that shown in Fig. 2 ) be voltage unstable if it includes at least one
utilising a standard EBP training algorithm, a unstable bus. Information such as voltage weak
comparison in performance was made between busedareas is becoming vitally important to
the two ANNs. In this respect, the latter ANN voltage stabilisation and control of modern
also had nine variables but, unlike the CUS power systems. As an attractive alternative to
network, each window length used had four the MLP and other conventional approaches to
samples; a smaller window length made the voltage stability analysis, KNs offer some
convergence of the network to the requisite advantages particularly in the application of
value extremely difficult. The hidden layer had clustering type problems.
18 neurons and the output vector again As mentioned before, a Kohonen self-
comprised seven variables. organising ANN is an unsupervised network
An extensive series of case studies showed which maps m-dimensional input vectors into
that the MLP-based network converged after two-dimensional outputs. Input patterns with
about 100 000 iterations (in approximately 45 similar features, which contain sufficient
minutes on a 133 MHz Pentium PC) and information about voltage stability of power
reached root-mean-square (RMS) error of 0.1. systems, are clustered together in an output
9 IEEE 30-bus system On the other hand, the CUS-based network map. The weight factors associated with the
output map neurons are then used to rank
buses from a viewpoint of voltage stability.
Thus power system conditions can be
effectively monitored in terms of voltage
instability, and effective preventive or
corrective control can be implemented to
enhance the overall system performance.
Example applications of the IEEE 30-bus
system are presented here to illustrate how KNs
can be applied to solving power system
problems. The system is shown in Fig. 9 and
includes six generation buses, 21 load buses
and 41 transmission lines.
Theorectical studies and utility experiences
all indicate that voltage collapse is mainly
attributed to heavy loading and/or system
contingencies. A comprehensive analysis of the
behaviour of power system variables under
voltage instability indicates that the following

284 POWER ENGINEERING JOURNAL DECEMBER 1998

 Tutorial: A N N s in power systems

two types of variables are important and can be

used as the input to the ANN: (i) load bus
voltage information, and (ii) reactive power
generation and reactive power loss. There are
thus a total of 23 inputs which include 21
giving load voltage information, one for total
reactive power generation and one for total
reactive power loss. It has been found that 12 x
12 neurons in the Kohonen layer give the best
performance for this particular system.
In order to determine training input vectors
of the KN, a load flow analysis of the system
was performed. In the training phase, a large
number of operating conditions relating to
voltage stability were considered. In this
study, 3500 input vectors were generated in
which 2000 were used for training the
network and 1500 for testing the network.
The operating conditions covered the range
from voltage stable states to a voltage unstable
state.
After training, the neurons of the KN are case and, in such cases, no voltage stability 10 Cluster in Kohonen
clustered. Fig. 10 shows the classification problem exists. Cluster B features the operating layer
results. Moreover, the properties of the clusters conditions in which the system faces severe
can be identified by analysing the weight voltage problems caused by load increas'es (for
vectors of the neurons in the clusters. Each example, an increase in load above 30% from
symbol in Fig. 9 represents a neuron of the the base load) or contingencies, and voltage
Kohonen layer. The neurons with the same control actions should be taken to improve
colour belong to the same cluster. There are stability Cluster C classifies intermediate cases
three clusters in the output map which between clusters A and B, in which some
represent three types of operating states in devices, such as generators, reactive
terms of voltage stability Cluster A groups all compensators and on-load tap changers, will
safe operating conditions created by load reach the operating limits. For a given cluster,
variation of -15% up to 17.5% from the base the properties of the cluster can be identilried by

11 3D picture of
weight vectors
corresponding to input
reactive power loss

POWER ENGINEERING JOURNAL DECEMBER 1998 285

 wey systems

12 Value of energy power flow analysis. position of the fired neuron indicates the degree
function against number From the viewpoint of voltage stability, the of voltage stability On the other hand, weight
of iterations
left top corner of the output map and its nearby vectors of the fired neuron associated with each
13 Generations of
neurons classify the weakest operating load bus input can be used to rank the buses.
units against number of conditions whereby the critical zone covers For example, the ranking of weak buses from
iterations more buses. Although there are no the weak to the strong (under base load
misclassifications between clusters A and B, conditions) with neuron 8 fired is: bus 30, bus
they do occur between clusters B and C 26, bus 29, bus 24, bus 23 etc.
whereby six vectors out of 80 cases belonging
to cluster B are classified into cluster C. Hopfield neural networks for solving power
For the trained KN, the properties of the system economic dispatch problems
clusters can be further explained by There are a large number of different
analysing the weight vectors of the neurons combinatorial optimisation problems facing
in the clusters corresponding to the reactive electricity utilities. The deregulation of the
power loss input. For a given cluster, the electricity supply industry worldwide adds
weight vectors of the corresponding neurons ever-growing motivations to develop new
represent the operating conditions with optimisation algorithms so as to design the
similar voltage stability. Fig. 11 shows the best strategies for most effectivelyutilising the
component values corresponding to total assets under increasing commercial pressures.
reactive power loss of weight vectors of each Various algorithmic and heuristic approaches
Kohonen neuron. It is apparent that the have been adopted or investigated by power
classification is consistent with the engineers. In particular, for the last few years
indication of total reactive power loss. For there has been a growing interest in
example, neuron 133 and its neighbours algorithms inspired by the observation of
classify voltage unstable operating natural phenomena to help solve complex
conditions, while neuron 12 and its computational problems. In this respect, HNs
neighbours classify voltage stable operating have been applied to power system optimisa-
conditions. However, the relationship is tion problems. Herein, the application in
nonlinear in nature. Note that the bottom solving a power system economic dispatch
wall of the three-dimensional picture stands problem is presented. The basic requirement
for a two-dimensional neuron grid and 144 of power economic dispatch is to generate
neurons are shown on it. adequate electricity to meet continuously
When a test vector is input into the trained changing customer load demand at the lowest
KN, one of the neurons in the Kohonen layer possible cost under a number of constraints.
will be fired. From the foregoing discussion, the Essentially, in applying HNs for solving

286 POWER ENGINEERING JOURNAL DECEMBER 1998

 Tutorial: ANN$ in power systems

power system optimisation problems, the

energy function of the HN has to be made
equivalent to a certain objective function that
needs to be minimised. The search for a
minimum energy performed by the HN
corresponds to the search for a solution to the
economic dispatch problem. The constraints
are mapped into the network by taking
feedbacks from variable neurons (computing
the optimisation variables) to the constraint
neurons (computing the constraint functions).
Thus, the key is to formulate the economic
dispatch problem with its objective function
which can be used to construct a HN, i.e. the
weights. Once the problem has been
formulated in this manner, a particular type of
parallel algorithm would have been
constructed. It is clear that the formulation is
application-oriented. For example, in a 15-unit
economic dispatch problem, Fig. 12 shows the
variation of energy function of the HN with ’r ... -ANN PSS
respect to the number of iterations. Fig. 13
describes the variation of generation level with
respect to the number of iterations.

Recurrent n e u r a l networks for designing

power system stabilisers
A major concern for power engineers is to keep
power systems under control at all times. This
is more so since power systems are being 0 1 2 3 4 5 6 7 8 9 10
pushed much closer to their stability limits for time, second
various reasons. In this respect, a number of
power system dynamic stabilisation techniques
have been employed by utilities. However, and hence the potential of ANNs in designing 14 Structure of the
there are still some difficulties/problems in new adaptive control for power systems. control system
designing more effective control strategies. In
recent years, various types of ANNs have been Parts 1 and 2 of this tutorial were uublished in the l 5 Response to a
June 1997 and February 1998 issue’s of the Journal. three-phase
employed for designing new power system circuit at the middle of a
stabilisers (PSS). The most commonly transmission line Cline
employed is the back-propagation network Further reading disconnected after
1 EL-SHARKAWI, M., and NEIBUR, D. (Eds.): o.05s and successfully
which can be generalised to the current ANN by ‘Artificialneural networks applied to power systems’, reclosed 3.93s later,
adding feedback connections. IEEE Power Engineering Society tutorial course,
Fig. 14, for example, shows the structure of IEEE catalogue number 96 TP 112-0, 1996
an adaptive power system stabiliser which has 2 ‘Intelligent system applications to power systems’,
two recurrent ANNs. One functions as a tracker Proc ISAP ’94, September 1994, Vols 1 and 2,
Montpelier, France
to learn the dynamic characteristics of the 3 WARWICK, K., EKWUE, A., and AGGARWAL, R.
power plant and the second functions as a (Eds.): ‘Artificial intelligence techniques in power
controller to dampen the oscillations caused by systems’ (IEE Power Engineering Series 22, 1997)
disturbances. In this approach, the weights of
the ANNs are updated online to consider any
0IEE: 1998
new information available during actual
control of the plant. A typical simulation result Raj Aggarwal is Reader in the Power and Energy Systems
Group, Department of Electronic & Electrical
is shown in Fig. 15 for a three-phase short-
Engineering, University of Bath; Yonghua Song is
circuit fault. The study clearly illustrates the Professor in the Department of Electrical Engineering &
superior performance attained over a Electronics, Brunel University. Dr. Agganval is an IEE
conventional power system stabiliser (CPSS) Fellow and Prof. Song is an IEE Member.

POWER ENGINEERING JOURNAL DECEMBER 1998 287