A CONCEPTION OF NEURAL NETWORKS IMPLEMEN
TATION
IN THE MODEL OF A SELF-LEARNING VIRTUAL POWER
PLANT
Robert Kucęba, Leszek Kiełtyka
Department of Management Information Systems, Management Faculty, Częstochowa University of Technology
Al. Armii Krajowej 19B, 42-200 Częstochowa, Poland
Keywords: virtual power plant, artificial intelligence, learning organization.
Abstract: The present article focuses on learning methods of self-learning organization (on the example of the virtual
power plant), using artificial intelligence. There was multi-module structure of the virtual power plant
model presented, in which there were automated chosen learning processes of the organization as well as
decision making processes.
1 INTRODUCTION
The article presents the virtual power plant model
(Kucęba, 2003), in which chosen decision making
processes are automated by application of neural
networks. The structure of the proposed model
consists of the following modules (Figure 1):
- The Internal Knowledge Components
Acquisition Module,
- The External Knowledge Components
Acquisition Module,
- The Explaining-Concluding Module,
- The Knowledge Aggregation Module,
- The Individual Units Motion Management
Module,
- The Control Module,
- The Heat Demand Prediction Module (for the
units producing heat in association).
It should be stressed that three of the above
mentioned modules (The Explaining-Concluding
Module, The Individual Units Motion Management
Module, The Heat Demand Prediction Module),
were worked out on the basis of artificial
intelligence (Kucęba, 2004). These modules realize
self-learning organization’s tasks, based on
identification of key competence essential to
coordinate and manage units in the virtual structures.
They were designed with the aid of the neural
networks professional simulator “Statistica Neural
Networks”.
2 FUNCTIONALITY OF THE
MODULES IN THE MODEL OF
THE VIRTUAL POWER PLANT
After defining structural modules, verification of the
acquired and generated information in the aspect of
their usefulness in the process of learning of the
simulated and determined in the project self-learning
virtual power plant was conducted.
The element aiding decision making processes in
the environment of dispersed production units of low
power is The Individual Units Motion Management
Module, based on neural networks. Input variables
set of this module is generated by the proceeding
Explaining-Concluding Module and The Internal
Knowledge Components Acquisition Module.
Neural networks implemented in The Individual
Units Motion Management Module generate
knowledge on the basis of which decisions in the
area of correlated units motion management are
made. Knowledge is then delivered to The Virtual
Power Plant Motion’s Central Dispatcher, which on
the basis of it and data acquired form The Data
Aggregation Module determines among other things
work schedule for the next hour or twenty four hours
of the virtual power plant and also manages motion
of the individual units producing low power energy.
Effective functioning of the proposed organization is
conditioned by proper selection of input variables in
the individual modules of knowledge components
acquisition. To this end there were two different
366
KucÄ
´
Zba R. and Kiełtyka L. (2005).
A CONCEPTION OF NEURAL NETWORKS IMPLEMENTATION IN THE MODEL OF A SELF-LEARNING VIRTUAL POWER PLANT.
In Proceedings of the Seventh International Conference on Enterprise Information Systems, pages 366-369
DOI: 10.5220/0002547003660369
Copyright
c
 SciTePress

methods of input variables selection used. In case of
The Internal Knowledge Components Acquisition
Module technical-economical analysis was used.
Verification of the input variables of The
Explaining-Concluding Module was based on
Crucial Factors Success Analysis. It was conducted
in the context of variables selection representing real
time changes occurring on the balance market and
the Energy Exchange. Effective management of the
individual units motion is largely dependent on
proper selection of knowledge components used in
the process of learning of neural networks allocated
in The Explaining-Concluding Module. In
connection with that conducted Crucial Factors
Success Analysis was limited to minimizing
function of errors generated in the structure of neural
networks using sensibility analysis, implemented
from applied in the research “Statistica Neural
Networks” – neural networks simulator. At this
stage of the conducted research for balance market
and the Energy Exchange there were independent
sets of neural networks of diverse typologies
defined, with the proper selection of input and
output variables.
The main feature of the applied sensitivity
analysis was determination of relative influence of
the input variables (knowledge components) on
neural network capacity by multiple network tests
and reduction of time series representing input
variables. The aim of this analysis was identifying
variables that can be omitted in the process of
learning without quality loss of neural network.
Applied sensitivity analysis evaluated variables
usefulness by giving them proper rank, ordering
observed variables according to their importance
(decreasing error). Selection determinant of the set
elements were values of prognosis errors generated
on the outputs of the networks.
As a result of the conducted analysis there was shape
of time series representing input vectors of neural
networks formulated. Data set structure, the
components of which are input vectors and assigned
to them input variables, was implied by specificity
of information aggregated and processed by
individual structural elements of self-learning virtual
Figure 1: Model of self-learning organization in the virtual management environment (power plant).
Source: own analysis
A CONCEPTION OF NEURAL NETWORKS IMPLEMENTATION IN THE MODEL OF A SELF-LEARNING
VIRTUAL POWER PLANT
367

organization (Duch, Korbicz, Rutkowski,
Tadeusiewicz, 2000). Data set of The Explaining-
Concluding Module was determined at this stage,
the components of which were vectors described by
attributes bunch.
In the modules – The Explaining-Concluding
Module and The Individual Units Motion
Management Module, individual attributes of the
vectors represent input and output variables
processed by carefully selected at the later stage of
research architecture of neural networks. In the first
stage of the project there were two independent data
sets for The Explaining-Concluding Module worked
out, in order to represent phenomena occurring on
the balance market and the Energy Exchange. Data
sets structure, on the basis of which the process of
learning of the presented module implemented in the
model of self-learning organization in the virtual
environment was implemented, was determined in
the following stages:
- Aggregation of information describing external
environment of a virtual energetic enterprise.
This information included variables such as:
energy price, its quality and defined within the
confines of research balance index in twenty-
four-hour/hour turn on the balance market and
the Energy Exchange;
- Division of data set into the following sub-sets:
learning, validating and testing (Kiełtyka,
Kucęba, Sokołowski, 2004). The structure of
these sub-sets was determined by constant
number of vectors’ attributes, which
represented values of input and output
variables.
3 INPUT AND OUTPUT
VARIABLES AGGREGATION
Selection of neural networks (Kiełtyka, Kucęba,
Sokołowski, 2004) was realized in a way separated
for information from the balance market and the
Energy Exchange. In case of the prediction process
on the Energy Exchange there were two complex
sets of diverse architecture neural networks
determined.
The structure of the learning, validating and
testing patterns, was defined here as two-
dimensional time series including the following
variables:
Input variables:
- Energy price on the Energy Exchange in twenty-
four-hour/hour turn (quantity variable),
- Energy amount on the Energy Exchange in twenty-
four-hour/hour turn (quantity variable),
Output variables:
- Energy amount (quantity variable),
- Energy price (quantity variable).
The variables were aggregated in time series the
range of which was determined from 5 to 10
elements (time series size). Prognosis horizon was
set an hour and twenty-four hours in advance.
Table 1 presents specification of the chosen neural
networks (MLP, GRNN, RBF) that constitute
executive element in the processes of energy price
prognosis (EP) and its amount (EA), implemented in
the Explaining-Concluding Module. There were
chosen days of observation in the year 2003
presented in the table. High quality of neural
networks in the process of prognosis (example table
1) determines assumption that the proposed element
of information processing increases functioning
efficiency of the simulated self-learning organization
in the virtual management environment.
At the next stage of data processing aggregated
knowledge components (omitting components form
the separated layer) were sent to the module that
constitutes the decision making aiding element of
the proposed model - The Individual Units Motion
Management Module. There were sets of neural
networks implemented in this module, which aided
classification of all the knowledge components in
order to generate information facilitating motion
management of the virtual power plant.
Input variables:
- costs of energy carriers acquisition in the
regional perspective (taking into consideration
renewable energy carriers), applied in the
production process of final energy,
- electric/heat indicative power of individual units,
- discretional power,
- unit price of the energy from scattered and
dispersed sources (heat and electric energy):
− unit price of fuel in the regional perspective,
− depreciation costs of energetic equipment,
− unit price of calculable interests,
− fuel value of individual energy carriers,
Table 1: Chosen neural networks typologies implemented
in the Explaining-Concluding Module in the process of
knowledge components prognosis.
ICEIS 2005 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
368

− energetic equipment efficiency,
- accounting prices of deviations prognosis (APD)
on the balance market in twenty-four-hour/hour
turn,
- energy amount prognosis on the balance market
in twenty-four-hour/hour turn,
- balance index prognosis in twenty-four-
hour/hour turn,
- energy price on the Energy Exchange prognosis
in twenty-four-hour/hour turn,
- energy amount prognosis on the Energy
Exchange in twenty-four-hour/hour turn,
- geographical location of individual correlated
units,
Output variable:
- dual state variable signaling turning individual
unit on/off (quality variable).
On the basis of key competence there were
defined priority criteria determining learning process
and self-learning of neural networks, which
determined proper work control of the correlated
production units. The following criteria were
formulated: type of energy sources used (renewable
sources having the highest rank), unit cost of energy,
demand for energy and geographical location. They
imply classification type of the processed knowledge
components, on the basis of which The Individual
Units Motion Management Module generated on
output dual state nominal variables, informing of
turning on/off individual units within self-learning
organization in the virtual management
environment. This information was passed to the
Virtual Power Plant Motion’s Central Dispatcher.
The dispatcher receives all data included in The
Knowledge Aggregation Module in the same time
unit. Access to data from both modules creates
situation where the dispatcher plays the role of
decision-maker managing effective functioning of
the virtual power plant. The dispatcher also
cooperates in the process of learning of The
Individual Units Motion Management Module,
verifying in real time generated by this module
information on the basis of all data components.
Decisions controlling motions of the individual
production units are passed to The Control Module
(switch), which is integrated with it by energetic
networks and VPN networks. In addition, VPN
networks realize communication among the
correlated units transferring internal knowledge
components (describing internal environment) to
The Internal Knowledge Components Acquisition
Module (Kiełtyka, Kucęba, 2003).
4 SUMMARY
The organization, for which the virtual management
environment was created, is integrated by the
network of mutual connections structure of
geographically dispersed low power energy sources.
Implementation of the designed framework
model may influence organizational efficiency
growth.
This article and its realization were inspired by
the research conducted by the author concerning
application of neural networks in various business
processes. In the author’s opinion creation of self-
learning organization in the virtual management
environment is not only the future but also the need
of the present day.
„The research financed from the funds of The
State Committee for Scientific Research in the years
2004-2006 as the research – 1H02D0727”
REFERENCES
Duch, W., Korbicz, J., Rutkowski, L., Tadeusiewicz, R.,
2000. Biocybernetics and Biomedical Engineering.
Neural Networks. Volume 6”. Ed. By Maciej Nałęcz,
Akademicka Oficyna Wydawnicza Exit, Warszawa
(in Polish).
Kiełtyka, L., Kucęba, R., 2003. Information and
Communication Technologies in the Virtual
Organizations. 5-th European Conference of Young
Research and Science Workers in Transport and
Telecommunications „TRANSCOM 2003”, Section 3,
Information Technologies Communication Systems
and their Control, Žilina, Slovak Republic 2003.
Published by University of Žilina.
Kiełtyka, L., Kucęba, R., Sokołowski, A., 2004.
Application of Neural Network Topologies in the
Intelligent Heat Use Prediction System. Proceedings
of the 7
th
International Conference Zakopane, Poland,
June 2004. Rutkowski L., Siekmann J., Tadeusiewicz
R., Zadeh L.A. – Artifical Intelligence and Soft
Computing – ICAISC 2004. Springer-Verlag Berlin
Heidelberg New York, pp. 1136-1141.
Kucęba, R., 2003. A virtual power plant basing on
dispersed and distributed generation. Proceedings of
the International Association for Development of the
Information Society International Conference
e-Society 2003. Volume I. Lisbon, Portugal.
Kucęba, R., 2004. Virtual Dimension of Open Electric
Energy Market. Proceedings of the 26
th
International
Conference on Information Technology Interfaces –
ITI 2004. Cavtat/Dubrovnik, SRCE University
Computing Centre, University of Zagreb, Croatia, pp.
27-28.
A CONCEPTION OF NEURAL NETWORKS IMPLEMENTATION IN THE MODEL OF A SELF-LEARNING
VIRTUAL POWER PLANT
369