Application of Convolutional Neural Network to PWR Fuel Loading Pattern Optimization by Simulated Annealing = 모의 담금질 방법을 이용한 가압경수로 핵연료 장전 모형 최적화에 합성곱 인공 신경망 적용

다국어 초록 (Multilingual Abstract)

Simulated Annealing (SA) algorithm is a widely used optimization method in the loading pattern (LP) optimization problems. However, the existing SA using neutronics code has a disadvantage of high computational cost. Therefore, the purpose of this paper is to reduce the computational cost of SA by applying prediction models using deep learning and Screening Technique. STREAM/RASTK 2.0 (ST/R2) code system was used to generate training data of prediction models and to perform 3D evaluation of SA.
As a simple evaluation code to replace the existing neutronics code, the prediction models of neutronic design parameters using deep learning was developed. As input parameters of the training data, assembly fuel enrichment, fraction of Burnable Poison (BP), number of BP rods, and assembly burnup were selected because they can be obtained from the specifications of assembly without additional calculation. In addition, through performance comparison, Convolutional Neural Network (CNN) was adopted as deep learning method, and prediction models was developed to calculate the peaking factor and the cycle length of the Korean Standard Nuclear Power Plant (OPR-1000).
The screening technique was developed to reduce computational cost by reducing the number of 3D evaluations in SA. The existing screening technique is a method of reducing computational cost through 2D evaluations instead of 3D evaluations using neutronics code. In this study, instead of 2D evaluations using neutronics code, the prediction models using deep learning that is faster than the existing code was applied as the simple evaluation code. By applying the prediction models to the screening technique, the computational cost can be further reduced. Also, uncertainty of the final LP due to the error of the prediction models can be reduced. For the design limit setting conditions, independent SA simulation with the screening technique were performed. As the results of the LP optimization using the simple evaluation code, the computational cost of optimization was greatly reduced.

목차 (Table of Contents)

• Contents
• Contents of Table ii
• Contents of Figure iii
• Abstract (English) iv
• 1. Introduction 1
• 1.1 Research background 1
• 1.2 Computer codes 3
• 1.3 Application of ANN 4
• 2. Prediction models of neutronic design parameters 8
• 2.1 Data generation 8
• 2.2 Deep learning methods 12
• 2.3 Performance comparison of deep learning method ( DNN vs. CNN ) 14
• 2.3.1 General models of DNN and CNN 15
• 2.3.2 Results of the performance comparison 16
• 2.4 Result 22
• 2.4.1 Architecture of CNN 22
• 2.4.2 Training result of the peaking factor 22
• 2.4.3 Training result of the cycle length 24
• 3. Loading pattern optimization 26
• 3.1 Simulated annealing algorithm 26
• 3.2 Screening technique 27
• 3.3 Application and result 28
• 3.3.1 Application (Shin-Kori unit 1) 28
• 3.3.2 Result 32
• 4. Conclusion 36
• References 38
• Abstract (Korean) 39