Abstract
A way to confine potential consequences of an accident in
a nuclear reactor is to limit the number of fuel rods
expected to fail in the course of the incident.
Particularly, the safety regulations in Finland require
that the number of failed fuel rods in the most severe
accident scenario would be less than 10% of all the rods.
In safety assessments, the estimation of the fraction of
failing rods is conventionally based on conservative
analyses, but this approach has several downsides.
Sometimes it is hard to judge whether the assumptions are
conservative because the phenomena in the reactor are
highly nonlinear. Additionally, conservative methods
often lead to excessive margins and that way to economic
losses. As a result, statistical best-estimate methods
have acquired an established position during the past two
decades. The development started worldwide when the
U.S.NRC revised its rules in 1988 to allow realistic
best-estimate methods complemented with uncertainty
analysis alongside with the old conservative approach.
These methods are based on the selection and variation of
parameters that are important in accident conditions. The
accident scenario is simulated with a designated computer
programme several times with different parameter values
between simulations, and that way an estimation of the
number of failed rods is obtained. In order to the
results to be statistically reliable, enormous number of
simulation runs is needed. Thus the analysis requires a
lot of computer resources, and this has been a limiting
factor for the breakthrough of this procedure. Different
approaches have been used to reduce the number of fuel
performance code calculations.
Before the current efforts in this field at VTT, there
has not been a statistical or any other systematic tool
in Finland for the evaluation of rod failures. Since
2006, a calculation system for statistical fuel failure
analysis has been under development. The calculation
procedure introduces neural networks as a new way to
reduce the number of simulations. Neural networks are
familiar from other applications in nuclear plant
modelling but the concept is a novelty in this context. A
neural network is trained with results from stacked fuel
performance code calculations, and then the network is
used as a substitute for the analysis code. Further, the
developed method utilizes the results of nonparametric
statistics, the Wilks' formula in particular. The system
has been successfully tested with a small scale analysis
and it is now ready for full reactor scale applications.
Original language | English |
---|---|
Title of host publication | SAFIR2010. The Finnish Research Programme on Safety 2007-2010. Final Report. Puska, Eija Karita & Suolanen, Vesa. VTT Research Notes 2571 |
Place of Publication | Espoo |
Publisher | VTT Technical Research Centre of Finland |
Pages | 132 - 144 |
ISBN (Print) | 978-951-38-7689-0 (soft back ed.), 978-951-38-7690-6 (URL: http://www.vtt.fi/publications/index.jsp |
Publication status | Published - 2011 |
MoE publication type | D2 Article in professional manuals or guides or professional information systems or text book material |