The subject of the research relates to the development and
implementation of algorithms for short-term prediction of the district
heating system characteristics using artificial neural networks. The
research is aimed at developing algorithms for the selection of
standard feedforward and recurrent artificial neural networks and their
architectures, choice and adjustment their parameters, choice and
definition of adequate inputs, modification of network architecture
and its adaptation to meet the demands imposed by the application of
artificial neural networks for short-term prediction of heat load as
main characteristic od district heating system. Special attention will
be devoted to a comparative analysis of proposed and adopted
artificial neural networks with their different architectures to obtain
optimal algorithms.
An adequate heat load prediction and satisfying consumer demands
with delivered heat energy in sense of control system, energy saving
and environment prot...ection, are very important preconditions for
optimal adjusting of district heating system
Improving quality of prediction, as one of the dissertation objective,
has positive impact to control of district heating system, in general.
The main focus is on adequate choice of input vector, number of input
nodes and other parameters for standard types of neural networks,
contrary to solutions of some authors from literature, where they are
creating totally new and unique networks for solving specific
problem. On that way, they are loosing possibility of generalization
which is opposite to one of the dissertation objective.
Specific attention is given to problem of transient regime of heating,
where there are no continuation in heating during a day and defined heating period.
Achieving qualitative prediction for short period is very important for
decrease heat consumption and increase the coefficient of equipment
exploitation. This is more important due the fact that district heating
systems in Serbia are intermitted by definition which means that
heating is not realized in continuation but with turning on and off in
the morning and evening hours. Short term prediction is realized for
prediction of selected parameters and district heating system
characteristics for period of one, three and seven days.
Deigned modified feedforward and recurrent neural networks satisfy
needed quality of prediction for district heating systems, adequately
predict peak loads in transient heating regimes and through the
realization of neural networks of the same architecture on four
different data heat sources, they are showing possibility of
generalization on specific level.
