Multi-Stage Neural Network Framework for Probabilistic-to-Continuous Predictions

James Daniell, Kazuma Kobayashi, Syed Alam, Ayodeji Alajo, Souvik Chakraborty, Dinesh Kumar

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Machine learning tools are becoming more popular for engineering problem solving, including those in nuclear systems. Nuclear reactor systems are complex, and cannot necessarily be analytically solved or simulated without a significant amount of resources. However, high reliability models with low computational resources are desired. A new type of model, Multi-Stage Deep Neural Network (MSDNN), is proposed to create a high-reliability machine learning model by using probabilistic classification information to generate more accurate continuous predictions. Multiple MSDNN architectures are designed and compared for predicting the final steady state power of a research nuclear reactor after a power changes. Results show that MSDNN models tend to perform better than standard feedforward neural networks in terms of both accuracy and generalization. This type of model is expected to be utilized and extrapolated for future nuclear problem solving using machine learning techniques.

Original languageEnglish (US)
Title of host publicationProceedings of 13th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies, NPIC and HMIT 2023
PublisherAmerican Nuclear Society
Pages1278-1287
Number of pages10
ISBN (Electronic)9780894487910
DOIs
StatePublished - 2023
Externally publishedYes
Event13th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies, NPIC and HMIT 2023 - Knoxville, United States
Duration: Jul 15 2023Jul 20 2023

Publication series

NameProceedings of 13th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies, NPIC and HMIT 2023

Conference

Conference13th Nuclear Plant Instrumentation, Control and Human-Machine Interface Technologies, NPIC and HMIT 2023
Country/TerritoryUnited States
CityKnoxville
Period7/15/237/20/23

Keywords

  • Machine Learning
  • Nuclear Reactor
  • Physics Informed
  • Predictive Modeling

ASJC Scopus subject areas

  • Human-Computer Interaction
  • Energy Engineering and Power Technology
  • Nuclear Energy and Engineering
  • Control and Systems Engineering

Fingerprint

Dive into the research topics of 'Multi-Stage Neural Network Framework for Probabilistic-to-Continuous Predictions'. Together they form a unique fingerprint.

Cite this