Multi-Stage Neural Network Architecture for Improving Continuous Prediction Reliability

Daniell, James; Kobayashi, Kazuma; Kumar, Dinesh; Chakraborty, Souvik; Alam, Syed Bahauddin

Permalink

https://hdl.handle.net/2142/121821 Copy

Description

Title

Multi-Stage Neural Network Architecture for Improving Continuous Prediction Reliability

Author(s)

• Daniell, James
• Kobayashi, Kazuma
• Kumar, Dinesh
• Chakraborty, Souvik
• Alam, Syed Bahauddin

Issue Date

2023

Keyword(s)

• Multi-stage deep neural network
• Reactor power
• Performance
• Prediction reliability

Abstract

Machine learning methods have become popular in the nuclear engineering sector due to their relatively good accuracy for the computational cost associated with prediction capabilities. Compared to traditional simulation, machine learning techniques typically sacrifice accuracy and reliability to drastically decrease computation time. To overcome the negative side of this tradeoff, modern techniques focus on new methods to improve the reliability of these models in the interest of developing these models so that they can be applied practically for nuclear problems and systems. To improve regression model reliability, a new neural network framework is developed by the team. This developed framework, a Multi-Stage Deep Neural Network (MSDNN), utilizes multiple prediction paths to generalize information in order to improve the accuracy and reliability of desired parameters. Neural networks have displayed powerful regression prediction capabilities on their own due to their ability to learn inherent nonlinearities in a dataset, although sufficiently complex problems or diverse datasets can be difficult to learn at high accuracy without data memorization. Built upon our previous work, the goal of the MSDNN is to utilize readily available and useful information to improve the prediction capabilities of the target output. Similar to transfer learning approach, this additional information can be used by allowing the MSDNN to implicitly learn qualities of the system or phenomena important to the prediction, which can be used to improve the physics. For the MSDNN proposed, a target prediction of steady-state reactor power after a power change is utilized. Nuclear reactors typically attempt to operate at steady state powers, although power changes may be required during operation. This is especially true of research reactors, which must change power to meet the designations of an experiment. However, since nuclear reactors are complex systems, continuous regression-type predictions can be unreliable outside long-term time series analysis. To improve the reliability of the target output, a reactor power classification scheme is developed as an additional output which can be used to determine the approximate final reactor power before making the regression prediction. By making a power classification first, the input data for the neural network can be generalized during training by attempting the classification. Architecturally, input layers are followed by a hidden layer, which is then the input to a classification output. The hidden layer inputs to another hidden layer as well, which is then used as the input to the target regression prediction. Information is generalized in the first hidden layer (by training for both classification and regression) and then further trained for the second hidden layer and output. The performance of the outlined MSDNN is compared to a traditional regression network for benchmarking training, accuracy, and reliability. Results show significantly improved regression accuracy using the MSDNN model, with a loss in classification accuracy compared to traditional classification models. Ultimately, MSDNNs could be used for regression modeling to help improve reliability.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121821

Owning Collections