Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting and forecasting reactivity and flux using long short-term memory models in pebble bed reactors during run-in

Ian Kolaja, Ludovic Jantzen, Tatiana Siaraferas, Massimiliano Fratoni

TL;DR

The paper presents a data-driven approach to predict and forecast reactivity, flux, and power in pebble bed reactors during run-in using long short-term memory networks trained on zone-based depletion data from the PEARLSim-Serpent framework. By reducing flux and power meshes with PCA and engineering discharge-pebble features, the authors achieve high predictive fidelity on unseen data (e.g., $R^2=0.9914$ for testing). The work also demonstrates forward forecasting capabilities and introduces an AI-guided run-in optimization loop that iteratively perturbs operating parameters to approach target states. This methodology enables better understanding of long-term reactor dynamics and offers a pathway to shorten running-in times while maintaining safety and fuel qualification constraints.

Abstract

Pebble bed reactor (PBR) operation presents unique advantages and challenges due to the ability to continuously change the fuel mixture and excess reactivity. Each operation parameter affects reactivity on a different timescale. For example, fuel insertion changes may take months to fully propagate, whereas control rod movements have immediate effects. In-core measurements are further limited by the high temperatures, intense neutron flux, and dynamic motion of the fuel bed. In this study, long short-term memory (LSTM) networks are trained to predict reactivity, flux profiles, and power profiles as functions of operating history and synthetic batch-level pebble measurements, such as discharge burnup distributions. The model's performance is evaluated using unseen temporal data, achieving an $R^2$ of 0.9914 on the testing set. The capability of the network to forecast reactivity responses to future operational changes is also examined, and its application for optimizing reactor running-in procedures is explored.

Predicting and forecasting reactivity and flux using long short-term memory models in pebble bed reactors during run-in

TL;DR

The paper presents a data-driven approach to predict and forecast reactivity, flux, and power in pebble bed reactors during run-in using long short-term memory networks trained on zone-based depletion data from the PEARLSim-Serpent framework. By reducing flux and power meshes with PCA and engineering discharge-pebble features, the authors achieve high predictive fidelity on unseen data (e.g., for testing). The work also demonstrates forward forecasting capabilities and introduces an AI-guided run-in optimization loop that iteratively perturbs operating parameters to approach target states. This methodology enables better understanding of long-term reactor dynamics and offers a pathway to shorten running-in times while maintaining safety and fuel qualification constraints.

Abstract

Pebble bed reactor (PBR) operation presents unique advantages and challenges due to the ability to continuously change the fuel mixture and excess reactivity. Each operation parameter affects reactivity on a different timescale. For example, fuel insertion changes may take months to fully propagate, whereas control rod movements have immediate effects. In-core measurements are further limited by the high temperatures, intense neutron flux, and dynamic motion of the fuel bed. In this study, long short-term memory (LSTM) networks are trained to predict reactivity, flux profiles, and power profiles as functions of operating history and synthetic batch-level pebble measurements, such as discharge burnup distributions. The model's performance is evaluated using unseen temporal data, achieving an of 0.9914 on the testing set. The capability of the network to forecast reactivity responses to future operational changes is also examined, and its application for optimizing reactor running-in procedures is explored.

Paper Structure

This paper contains 16 sections, 5 equations, 15 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Data Generation
  3. Zone Based Core Simulation with PEARLSim
  4. PEARLSim Output and ML Feature Design
  5. Flux and Power Mesh Dimension Reduction
  6. Sequence Generation with PEARLSim
  7. Model Training and Assessment
  8. LSTM Training and Forecasting
  9. Feature Importance
  10. Model Performance
  11. Mesh Reconstruction
  12. Forecasting
  13. Run-in Optimization
  14. Algorithm Overview
  15. Performance
  16. ...and 1 more sections

Figures (15)

  • Figure 1: Serpent-generated plot of the gFHR model implemented in Serpent, showing the full core (top) and fuel pebble (bottom).
  • Figure 2: Schematic of core simulator zones, as well as the flow of pebbles through them. This schematic only has 3 radial zones and 4 axial zones.
  • Figure 3: Examples of the 20 by 8 subzone meshes generated by PEARLSim for power (top left), thermal flux (top right), epithermal flux (bottom left), and fast flux (bottom right). The power mesh typically has a relative uncertainty of 1.8%, while the flux mesh has 1.3%.
  • Figure 4: Demonstration of how last-pass burnup would be averaged for radial zones using HxF equilibrium data as an example. This shows how pebble-wise measurements can be spatially binned to inform the model.
  • Figure 5: Example of discharge pebble burnup binning for a core at low power early in operation (top) and at full power close to equilibrium (bottom).
  • ...and 10 more figures