voltage summing function.

Simulations on a 6-bus-4-generator system show that

neural network controllers are trained to provide better voltage regulation than
conventional controllers and previously unstable systems can be stabilized by a
properly trained neural network controller. This is quite a remarkable finding. It is
also shown that optimal recovery after a major disturbance is achieved by training
neural network controllers to provide optima! regulation for non-catastrophic faults
and also to provide stabilization for sorne catastrophic faults.
In the paper, optima! control of nonlinear systems using the open-loop dynamic
optimization algorithm, as well as the neural network state-feedback approach trained
by the backpropagation-through-time algorithm are reviewed. Comparing the
adaptation equations of the two, it can be seen that backpropagation-through-time and
dynamic optimization are related in the sense that a neural network can be trained to
give results that are very similar to those that dynamic optimization would give. In
other words, neural networks can give dynamically optima! solutions.
This is fortuitous, as dynamic optimization is not a practical algorithm for on-line
applications. Runs could take a few hours, depending on the complexity of the plant
and the computer speed. Even the fastest, most advanced computers still would not
make dynamic optimization practica!. On the other hand, neural networks can be
trained off-line, when time is not crucial. After adequate training, the weights of the
neural network can be loaded and used on-line. Response speeds would be in a matter
of milliseconds. However, dynamic optimization serves as a good benchmark for
determining the goodness and optimality of a trained neural network. Isolated cases
111
IIPI
could run with dynamic optimization, and the solution can be compared with the
neural network solution as a test.
The paper by Vasconcelos et al. [20] presents the capabilities provided by Kernel
Regression Trees - a hybrid non-parametric regression technique - to on-line dynamic
., security assessment and monitoring of isolated power systems with high penetration
of wind power. In the applied technique, to avoid overfitting a pruning algorithm is
used to extract the security structure. This approach, is demonstrated on the electrical
11111'
power system of Crete Island, and proves very suitable to extract simple,
interpretable, and reliable security structures. A description of the security problem
and the data set generation procedure are included. Comparative results regarding
performances of Regression Trees and Decision Trees are presented and discussed.
The security evaluation structures provided by the Kernel Regression Trees
approach have been integrated into CARE [ 17], the advanced control system that aims
to achieve optima! utilization of renewable energy sources, in a wide variety of
medium and large size isolated systems. The security evaluation structures obtained
provide a classification on dynamic security. Moreover, the degree of security is
provided, expressed as the expected minimum value of system frequency and
maximal rate of frequency change for a selected disturbance. For the creation of the
data set, a large number of initial operating points (OPs) were obtained by varying
randomly the load for each load busbar, the wind power for each wind park and the
wind margin. For each of the produced OPs a number of possible disturbances has
been simulated.
From the results obtained with the various approaches the following conclusions
are derived:

1
rt I. The Kernel Regression Tree approach is able to provide security classification
results and emulation of the numerical security index fm;• in a coherent way and
with good accuracy. Besides, KRTs provide simple interpretable security rules that
can be adopted by operators in the control rooms to help them in operating the
system.
I. Decision Trees (DT) provide classification structures of comparable performance
but with a simpler structure, which makes easier any interpretation of the
phenomena and of the influence of the relevant parameters. KRTs however have
the advantage of producing simultaneously a classification structure and giving the
degree of robustness of the system.
The paper by Matos et al. [12] tackles the same problem as the previous paper with
a different technique. lt reports on the application of fuzzy reasoning to the fast
assessment of the dynamic security of isolated power systems with high wind power
penetration. The inference method is a Takagi-Sugeno type system with a small
number of rules, optimized for each specific leaming set by a standard method
included in the MATLAB Fuzzy Logic Toolbox. The methodology is demonstrated in
a contingency study in the network of Crete that showed interesting results in the test
set. The paper discusses sorne implementation issues and possible future
developments of the approach, now in a preliminary phase.
Two approaches have been used regarding the classification problem: direct
classification of the operating state, or inference of the value of sorne index or
important variable, then used for classification. Sorne of the techniques can work with
either philosophy. In the present case, frequency is considered as the most important
variable, and both Ím;. and df/dt..., values constitute usual security indices that lead
straightforward to decision rules based on thresholds to their values. In this paper,
both approaches are used and results are compared.
Most of the paper is devoted to the analysis of the example and results. Results of
the application of fuzzy inference systems to dynamic security assessment are very
promising, even when a general purpose package is used to design the system.
However, more extensive tests are certainly needed, with different learning sets and
contingency situations, in order to draw more definite conclusions. Future
development of this work includes the use of different types of fuzzy inference
systems (first-order Takagi-Sugeno, different logical operators, etc.), new training
algorithms and new training philosophies, namely for the direct classification
procedures, with minimization of the classification error instead of the RMSE.
A Genetic Algorithm (GA) solution to the Economic Dispatch (ED) problem is
presented in the paper by Saramourtsis et al. [15]. Economic dispatch analysis
schedules the outputs of the online generating units so as to meet the system load at
mínimum cost. Improvements in scheduling the unit outputs can lead to significant
cost savings. Traditional dispatch algorithms (i.e. lambda iteration) are based on the
concept of equal incremental cost: the total production cost of a set of generators is
minimized when ali the units operate at the same incremental cost. When considering
transmission losses, the unit incremental costs are modified to account for incremental
transmission losses. Traditional algorithms however, require that the unit cost curves
($/h vs MW) be convex functions. Hence, they cannot handle non-monotonically
increasing incremental cost curves.
 Solution to the economic dispatch problem with non-convex unit cost functions
can be achieved using dynamic programming (DP). Unlike the traditional solution,
the DP solution imposes no restrictions on the generating unit characteristics.
However, it suffers from the dimensionality problem: as the number of generators to
be dispatched increases and higher solution accuracy is needed, the storage
requirements and the execution time increases dramatically.
The GA solution proposed in this paper is integrated in CARE and applied to Crete
with very satisfactory results. A binary genetic algorithm was implemented in arder to
achieve the operating cost minimization. The algorithm's aim is to maximize the GA
objective function (which is the inverse of the operating cost) while satisfying the
constraint relations that are set by the generators' operating limits, as well as by the
power demand. GA solutions do not impose any convexity restrictions on the
generators' cost functions making them ideal for use with generator units non-convex
cost functions. Test cases with different numbers of online generators showed that the
proposed GA gives comparable results with the well-known o...iteration method that
was used as benchmark. In addition the encoding strategy, the penalty policy for
constraints violation and the importance of special operators that were designed
specifically for the economic dispatch problem are analyzed.
The paper by Gavoyiannis et al. [8] addresses the problem of computing the
probability density functions (pdfs) of power flows as approximated by a weighted
sum of a finite number of Gaussian kernels. The parameters and weights of these
, '
kernels are iteratively estimated from the actual power flow input samples using the
EM (Expectation-Maximization) algorithm implemented by a suitable Neural
Network. In this way, probabilistic load flows can be calculated avoiding assumptions
about the input data regarding their pdfs and correlation among them. The proposed
method, still at its infancy, presents interesting prospects for planning applications.

2.2 Machine Learning Applications at the Power System Component Level

The paper by Kalles et al. [ 1 O] propases a novel approach to intelligent monitoring

and predictive maintenance of power plants based on reinforcement learning. This is a
new computational approach to automating goal-directed learning and decision­
making emphasizing on learning from direct interaction with its environment without
relying on exemplary supervision or complete models of its environment.
Proper maintenance of power plant equipment is essential to high system
availability. The complexity of modern power plant equipment is such, that few plant
maintenance personnel are truly experts in diagnosing faults and even fewer are
capable of predicting them. An effective approach to equipment maintenance is
predictive maintenance, and has been used to monitor the health of rotating
equipment by recording the vibrations and temperatures of their bearings. These
measurements, or signatures, are taken periodically, analysed using well-known DSP
techniques and compared with desired signatures or with earlier signatures to observe
how they evolve. When significant deviations occur, these are flagged so that
appropriate action can be taken. The decision on whether or not the equipment
requires maintenance is taken by human experts though expert systems are
commercially available to this end. Maintenance is carried out only when necessary
and not at regular intervals as in preventive maintenance, thereby reducing
maintenance costs and the possibility of majar breakdown significantly.
 Essential to predictive maintenance is any platform capable of data acquisition
from suitable sensors, which supply a continuous flow of data on the condition of the
plant in addition to a real-time decision mechanism to �re�ict impending faults so t�at
_
immediate remedia! action can be taken. Both qual1tat1ve as well as quant1tat1ve
information from the plant can be fused in the inference mechanism to decide on the
condition of the plant equipment and what maintenance is required and most
importantly, when. The diagnostic problem starts with the observation of sorne
deviation in the behavior of the plant equipment as compared to that desired or
expected. When a malfunction is observed, maintenance personnel hypothesize on its
causes, based on their knowledge and experience. Attempting to accept or reject sorne
of these hypotheses leads to further tests or uses specific or deep knowledge of the
specific piece of malfunctioning equipment. This knowledge usually concems the
structural and behavioral characteristics of the equipment. Qualitative reasoning has
attracted much attention in the past. Rule-based diagnostic systems with deterministic
or fuzzy reasoning have been proved capable of reaching conclusions using only
shallow knowledge about the plant. Artificial Neural Networks, by their very nature,
allow for leaming by viewing this problem as a pattem classification activity. Their
mapping ability can be exploited to perform associations between input patterns
derived from plant sensors and pattems representing fault conditions. Input pattems
are represented by vectors of measurements while the output data from the system are
vectors of the fault space representing the condition of the health of power plant
equipment.
Toe paper by Damousis et al. [4] presents a genetic fuzzy model that utilizes
special features, for the prediction of the wind speed variations in various time
windows in the future. Toe model utilizes a set of meteorological stations that encircle
the wind turbine clusters at a radius of 15km or more. The system has been applied
using data collected over a period of 2 years at locations in the Crete Island and
Northem Greece. A user intervention to assist the training process is not needed while
there is also no need for parameter initialization, like in other methods, contributing to
the model' s robustness.
Studies have also been made for spatial correlation of wind turbulence regarding
short distances (700 m to 15 km) and short time scales (wind changes per 4, 10, 30
min and also 1-minute deviations from 30-min averaged value). Correlation
coefficients are concluded to be related to the wind direction, terrain roughness and
height above the ground.
The paper induces a genetic fuzzy system that uses local and spatial relations of the
wind speed so as to improve the efficiency of forecasting, ranging from minutes to
severa! hours ahead. Since the efficiency of a genetic algorithm based solution
depends greatly on the coding scheme, the proposed GA for the training of the fuzzy
system uses an effective coding scheme as well as special operators, tailored to the
problem. The model is flexible and self-sufficient so as to easily leam and adapt in
data of any special case without any guidance or constraint by the user.
The paper by Georgilakis et al. [9] uses machine leaming to predict iron losses in
the transformer manufacturing process. Prediction of iron losses is an important task
in the transformer manufacturing industry, since iron losses constitute one of the main
parameters of transformer quality. Furthermore, accurate prediction of transformer
iron losses protects the manufacturer from paying loss penalties. In order to avoid this
risk, and in view of the fact that iron losses cannot be accurately predicted in the
current practice, one possible method is to design the transformer at a lower magnetic
induction, resulting in an increase of the transformer cost since more magnetic
material is required. Satisfactory prediction of iron losses, however, can be achieved
only if various parameters involved in the process, both qualitative and quantitative,
are taken into consideration. Instead, in the current practice, only the loss curve is
used, i.e., the influence of the rated magnetic induction on iron losses for each
specific magnetic material. This is dictated by the fact that there is no analytical
relationship expressing the effect of the other parameters on transformer iron losses.
The origin of this paper is the effective use of measurements taken at the first
stages of transformer construction, in order to minimize iron losses of transformer.
These measurements are collected and stored in databases. Each database corresponds
to different conditions (environment), i.e. to a certain supplier, grade and thickness of
magnetic material. When a satisfactory number of measurements has been collected,
machine learning is applied in order to learn the information included in the
databases. More specifically, decision trees are used to select the most relevant
attributes among a large set of candidate ones and to produce "if-then-else" decision
rules. These rules are applicable at the early stages of core production, and allow
possible corrective actions during the manufacturing process.
Moreover, neural networks are used to predict iron losses at the early stages of
transformer manufacturing. Each neural network model is suited to a different
environment. Selection of the most appropriate network (or equivalently
environment) is based on the satisfaction of customers' requirements and severa!
technical and economical criteria. The attributes selected by the decision trees are
used as inputs to the neural networks. The intelligent iron loss model (i.e., the model
of iron losses obtained through the neural network) is applied on-line in order to
optimally combine the individual cores and reduce the iron losses of assembled
transformers.
Finally, based on the good performance of the neural network structure, a new
grouping algorithm has been proposed. This algorithm selects, among the various
possible combinations of grouping cores, the one providing the minimal deviation of
predicting iron losses from the guaranteed to the customer losses. Application of the
proposed method to transformer manufacturing industry has verified the accurate
prediction of iron losses in ali the examined environments. Moreover, reduction of the
transformer losses is achieved.
Triantafyllidis [19] presented a method for prediciting mesh density with the use of
Artificial Neural Networks (ANN). The Finite Element Method (FEM) has been
widely used to solve electromagnetic field problems, mainly due to its ability to
handle cases of geometrical complexity. One of the main disadvantages of the method
is the need for an experienced user to provide a quality mesh, i.e. a mesh with which
the method will converge to an accurate solution quickly. One way around this
problem is the adaptive use, i.e. starting with an initial coarse mesh the problem is
solved, then the solution error is estimated for each element and finally the elements
with error that exceeds a given threshold are split into smaller elements, thus refining
the initial mesh. This technique will provide very accurate results, but is very time
and memory consuming.
The use of Artificial Neural Networks (ANNs) has been proposed for predicting
the mesh density of specific electromagnetic problems [2]. For this purpose a
prototype mesh is generated for a magnetic device, which includes ali features
expected to be found in future devices to be meshed. The device is sampled at specific
sample points and a set of geometric features is associated to the mesh density at the
neighborhood of the given sample point. The data gathered constitute the training
database for the ANN. Although these mesh density prediction methods work well for
the magnetic devices that they were designed for, they fail in cases of faulted
overhead power transmission line systems (TLS), carrying a zero-sequence current.
The paper presents an ANN trained to reproduce the mesh density vector of open­
boundary faulted power transmission line problems under the presence of small
features, such as conductors. In order to produce the data for the training database of
the ANN a technique for recursively subdividing the solving area is proposed. A
Delaunay-based mesh generator is used to produce the final mesh. Severa! test cases
show that this method significantly reduces solving time.
The paper by Satsios et al. [ 16] presents a novel application of genetically evolved
Fuzzy Logic Systems (FLSs), to transmission line electromagnetic field modeling. A
sophisticated artificial intelligence system (AIS), consisting of a sequence of FLSs
which evolve until the optimum FLS is achieved, has been developed by the
combination of FLS and GA techniques. The artificial intelligence system (AIS)
proposed uses a rule base adaptation mechanism. Instead of using fixed-length
chromosomes, new rules are added when necessary to form a variable-length FLS­
chromosome.
The AIS presented consists of a population of fuzzy logic systems (FLSs) evolving
with time. The evolution of the FLSs is accomplished by using a rule base adaptation
mechanism combined with a genetic algorithm (GA). The number and parameters of
fuzzy rules of each FLS are changed, until an optimum is achieved. To implement
this, a vector of variable length is defined for every FLS, containing its parameters.
This vector corresponds to an FLS-chromosome that will be processed by the GA.
After sorne generations, the population of the FLS-chromosomes converges to the
optimum FLS. The developed AIS is used to predict the electromagnetic field induced
by an overhead transmission line. This field depends on multiple variables, such as
the position of the phase conductors, the currents flowing through all conducting
materials, the earth resistivity and other operational parameters. The AIS was trained
using data resulting from finite element method calculations, for different
configuration cases of the above electromagnetic field problem. The performance of
the optimum FLS, derived after the training, is tested for new configuration cases,
differing significantly from the cases used for training. It is shown that the proposed
method may be very effective in predicting electromagnetic fields.
The architecture of each FLS of the AIS has been designed to salve the inductive
interaction problem between a faulted overhead transmission line and a nearby-buried
pipeline. This problem is of growing practica! interest, due to restrictions currently
imposed on public utilities in the use of right-of-ways resulting in overhead
transmission lines, pipelines, railroads, telecommunication lines etc. to be laid in
narrow corridors for severa! kilometers.
Fountas and Hatziargyriou [7] proposed the use of Petri Nets for planning power
system restoration. Experience from the development of system restoration plans
based on knowledge bases has indicated the effective role of these systems to the
overall restoration process. Recognizing the maturing capabilities of the expert
systems, it is believed that the suitable integration with existing facilities will enable
many elements of restoration planning to be carried out in an on-line environment.
This environment shall provide powerful on-Iine assistance to operators of the type
they most need, at the time they most need it. The result will be a significant
reduction in the time required for major restoration incidents, and significant
reduction in the losses that are incurred.
In this paper Petri Nets are used to plan power system restoration. Petri Nets
constitute a reasoning tool that has been used in modeling discrete event dynamic
systems, mainly in the areas of computers, communication protocols and factory
automation. This model-based approach is based on Petri Net fonnulation o[ the
power system procedures. Illustrative examples are provided, showing the
applicability of Petri nets in power system restoration (PSR) following a total or
partial blackout. Common features regarding power system behavior are highlighted,
and the equivalent models are explained.
By implementing the state-space analysis techniques, ali important net properties
can be analytically explained and verified. Through the analysis of the 1989 Hellenic
power system blackstart, the dynamic behaviour of the modelled power system was
evaluated and the suggested sequence of generic restoration tasks proved to be
meaningful and well-defined. The cumulative required restoration time was estimated,
assigning practica! time delays to transition firings.

3 Conclusions

In the Workshop a wide variety of ML techniques for solving timely problems in the
areas of Generation, Transmission and Distribution of modern Electric Energy
Systems have been proposed. In particular, applications of Artificial Neural
Networks, Kernel Regression and Decision Trees, Fuzzy Systems and Genetic
Algorithms have been proposed or applied to security assessment, control, economic
dispatch, power flows calculations, forecasting, restoration, plant monitoring,
transformer manufacturing, faulted transmission line problems, and inductive
inference problems. In ali cases, the application of ML techniques has proven their
efficiency to tackle difficult power system operation, planning and manufacturing
issues. These applications, by no means exhaustive, are indicative of a booming
activity in Power System applied research, as shown by the impressive number of
related publications in the literature. In conclusion, ML techniques coupled with
mature analytical techniques promise to provide significant support to power systems
in the new era.

References

1. "CARE: Advanced Control Advice for power systems with large scale integration of
Renewable Energy sources", contract JOR3-CT96-0119, Proceedings of CARE
Workshop, Heraklion 16-17 July 1999 and Final Project Report, September 1999.
2. Chedid and Najar, "Automatic Finite-Element Mesh Generation Using Artificial Neural
Networks - Part I: Prediction of Mesh Density," IEEE Trans. on Magnetics, pp. 5173-
5178, Vol. 32, n. 5, Sep. 1996.
• 3.

4.
IGRE TF38-06-06, "Artificial Neural Networks for Power Systems: A Literature Survey•·,
Engineering Intelligent Systems, Vol. l, No. 3, pp. 133-158, Dec. 1993.
I.G. Damousis, P.D. Dokopoulos: A Genetic Fuzzy Model for Wind Speed Prediction and
Power Generation in Wind Parks. In [22].
5. T.E. Dy Liacco, "Control of power systems via the multi-level concept", PhD Thesis, Case
Western Reserve University, 1968.
6. M. El-Sharkawi, D. Niebur (editors). "Artificial Neural Networks with Applications to
Power Systems", IEEE PES, 96 TP 112-0.
7. N. Fountas, N. Hatziargyriou: Artificial Intelligence Approaches to the Problem of Power
System Restoration. In [22].
8. A.E. Gavoyiannis, N. Vlassis, N.O. Hatziargyriou: Probabilistic Neural Networks for
Power Flow Analysis. In [22].
9. P. Georgilakis, N. Hatziargyriou: Machine Learning Applications in the Transformer
Manufacturing Industry. In [22].
10. D. Kalles, A. Stathaki, R.E. King: Intelligent Monitoring and Maintenance of Power
Plants. In [22].
11. G. Kariniotakis, G.S. Stavrakakis, E.F. Nogaret, "Wind Power Forecasting using advanced
neural network models", IEEE Trans. on Energy Conversion, 96 SM 552-0 EC, pp. 762-
767, Vol. 11, No. 4, Dec. 1996.
12. M. Matos, J.A. Pecas Lopes, M.H. Vasconcelos: Dynarnic Security Assessment by Fuzzy
Inference. In [22].
13. C.K. Pang, F.S. Prabhakara, A.H. El-Abiad, A.J. Koivo, Security evaluation in power
systems using pattern recognition, IEEE Trans. 011 PAS 93, no. 3, 1974.
14. A.O. Papalexopoulos, S. hao, T-M. Peng, "An Implementation of a Neural Network based
Load Forecasting Model for the EMS", IEEE Trans. on Power Syst., pp. 1956-1962, Vol.
9, Nr. 4, Nov. 1994.
15. A. Saramourtsis, J. Damousis, A. Bakirtzis, P. Dokopoulos: Genetic Algorithm Solution to
the Econornic Dispatch Problem - Application to the Electrical Power Grid of Crete
Island. In [22].
16. K.J. Satsios, 1.G. Damousis, O.P. Lambridis, P.S. Dokopoulos: An Innovative Genetic
Fuzzy System for the Prediction of an Inductive Interference Problem. In [22].
17. G.B. Sheble, K. Brittig, "Refined Genetic Algorithm - Econornic Dispatch Example", 94
WM 199-0, pp. 117-124, IEEE Trans. on Power Systems, Vol. 10, No. 1, Febr. 1995.
18. D.J. Sobajic, R. Shen, B. Widrow: Supplementary Neural Controls for Improved Transient
Behavior of Power Systems. In [22).
19. D. Triantafyllidis: Mesh Density Prediction for Open Boundary Faulted Power
Transrnission Line Problems Using Artificial Neural Networks. In [22).
20. M.H. Vasconcelos, J.A. Pecas Lopes: Pruning Kernel Regression Trees for Security
Assessment of the Crete Network. In [22].
21. L. Wehenkel, "Automatic Learning Techniques in Power Systems", Kluwer Acadernic
Pub!., 1998.
22. Proceedings of the Workshop on Machine Learning Applications to Power Systems,
Advanced Course on Artificial Intelligence (ACAI '99), Chania, Greece, 1999
(http://www.iit.demokritos.gr/skel/eetn/acai99/Workshops.htm).