The aim of this collaboration, between the division of Thermal Power Engineering
and Lunds Energi AB, is to investigate the possibilities of training artificial neural
networks (ANNs) with power plant operational data. For this purpose operational
data from Lunds Energi's gas turbine GT10B with heat recovery unit (HRU) will be
used. Furthermore a model with user interface is created to demonstrate the
possibilities of using ANN. The results are evaluated through feedback from Lunds
Energi and many different areas of implementation are considered.
ANN differs from conventional mathematical models in the sense that they are
trained rather than programmed. During training, data is presented in an iterative
manner in order to find the relation between selected inputs and outputs. After the
network is trained the weights, i.e. the parameters containing the network
information, are locked and if the network is presented with new, before unseen, data
it is able to predict new outputs.
The software used for modeling ANNs is called NeuroSolutions. The resulting
networks have been processed in Visual Basic for final use in Excel.
Thru these studies several ANN models have been produced, both models of the gas
turbine (GT) and models for sensor validation (SV). The results have been
promising, e.g. with networks demonstrating high performance predictability.
[1]
Jaime Arriagada.
On the Analysis and Fault-Diagnosis Tools for Small-Scale Heat and Power Plants
,
2003
.
[2]
Geoffrey E. Hinton,et al.
Learning internal representations by error propagation
,
1986
.
[3]
V. Kvasnicka,et al.
Neural and Adaptive Systems: Fundamentals Through Simulations
,
2001,
IEEE Trans. Neural Networks.
[4]
Mohsen Assadi,et al.
Fault prediction/diagnosis and sensor validation technique for a steam power plant
,
2005
.
[5]
Ihor S. Diakunchak.
Performance Deterioration in Industrial Gas Turbines
,
1991
.
[6]
Patrick van der Smagt,et al.
Introduction to neural networks
,
1995,
The Lancet.