Darling, Michael C., Jones, Thomas B., Luger, George F., Groth, Katrina M. ALADDIN: The Automatic Loader of Accident Data for Dynamic Inferencing Networks (Technical Report) Sandia National Laboratories, Albuquerque, NM, (SAND2015-10401), 2015.

BibTeX

@techreport{SAND2015-10401,
title = {ALADDIN: The Automatic Loader of Accident Data for Dynamic Inferencing Networks},
author = {Michael C Darling and Thomas B Jones and George F Luger and Katrina M Groth},
year = {2015},
date = {2015-11-01},
number = {SAND2015-10401},
address = {Albuquerque, NM},
institution = {Sandia National Laboratories},
abstract = {Sandia has initiated the development of a methodology to generate “SMART (Safely Managing nAccidental Reactor Transients) Procedures” to support nuclear power plant operators in diagnosis of severe accidents. The theoretical framework for developing SMART procedures involves coupling dynamic PRA, nuclear reactor simulation codes, and Bayesian Networks (BNs) provided by the University of Pittsburgh’s Structural Modeling, Inference, and Learning Engine (SMILE) and its graphical inferencing tool (GeNIe) to provide fast-running diagnostic support.

A critical challenge when dealing with the output from dynamic PRA simulations is that each accident simulation produces hundreds of megabytes of data over thousands of time steps. Only a fraction of this information can be included in models designed for faster-than-accident-time use. In order for the SMART Procedures concept to be scalable to nontrivial diagnosis scenarios, it became essential to develop a method to automate the handling of the inputs to the BN. Therefore, Sandia partnered with the University of New Mexico to develop a system that would automate the processing of simulation data for input into a BN model. This report outlines the development of the Automatic Loader of Accident Data for a Dynamic Inferencing Network (ALADDIN). ALADDIN automates the process of parsing and analyzing reactor simulation data to be input into GeNIe BNs. This was accomplished by creating a Java-based program which processes and discretizes the raw simulation data, calculates probability tables for the states of the plant parameters conditioned on the reactor states, and uses these calculations to build a bayesian network using the SMILE framework. Additionally, the system calculates the relative information gain value of each plant parameter. These calculations provide insight into which parameters would be most pertinent during particular accident scenarios. The initial prototype of ALADDIN outputs a diagnostic BN model for two accident types and twelve reactor parameters in a sodium fast reactor design.},
keywords = {Artificial intelligence, Bayesian Networks, Decision support systems, Dynamic PRA, nuclear power, software, uncertainty},
pubstate = {published},
tppubtype = {techreport}
}

Abstract

Sandia has initiated the development of a methodology to generate “SMART (Safely Managing nAccidental Reactor Transients) Procedures” to support nuclear power plant operators in diagnosis of severe accidents. The theoretical framework for developing SMART procedures involves coupling dynamic PRA, nuclear reactor simulation codes, and Bayesian Networks (BNs) provided by the University of Pittsburgh’s Structural Modeling, Inference, and Learning Engine (SMILE) and its graphical inferencing tool (GeNIe) to provide fast-running diagnostic support.

A critical challenge when dealing with the output from dynamic PRA simulations is that each accident simulation produces hundreds of megabytes of data over thousands of time steps. Only a fraction of this information can be included in models designed for faster-than-accident-time use. In order for the SMART Procedures concept to be scalable to nontrivial diagnosis scenarios, it became essential to develop a method to automate the handling of the inputs to the BN. Therefore, Sandia partnered with the University of New Mexico to develop a system that would automate the processing of simulation data for input into a BN model. This report outlines the development of the Automatic Loader of Accident Data for a Dynamic Inferencing Network (ALADDIN). ALADDIN automates the process of parsing and analyzing reactor simulation data to be input into GeNIe BNs. This was accomplished by creating a Java-based program which processes and discretizes the raw simulation data, calculates probability tables for the states of the plant parameters conditioned on the reactor states, and uses these calculations to build a bayesian network using the SMILE framework. Additionally, the system calculates the relative information gain value of each plant parameter. These calculations provide insight into which parameters would be most pertinent during particular accident scenarios. The initial prototype of ALADDIN outputs a diagnostic BN model for two accident types and twelve reactor parameters in a sodium fast reactor design.