Static Spatial Correction Factor Prediction Method Based on Convolutional Neural Network

Dynamic rod worth measurement (DRWM) stands as the predominant method for control rod worth measurement in low-power physics experiment of commercial pressurized water reactors (PWRs) worldwide, significantly reducing experiment timelines, eliminating borated wastewater, and improving nuclear plant economics. The static spatial correction factor (SSF) calculation is a critical component of this process, where SSF accuracy directly determines the reliability of rod worth measurements. Traditional SSF computation methods, such as Monte Carlo codes and deterministic codes, face challenges in computational complexity, massive data demands, and inefficiency. This study addresses these limitations by integrating deep learning to enhance SSF calculation efficiency and precision. Focusing on the Cycle 2 core of a domestic M310 reactor, an in-depth analysis of refueling design rules was conducted. By refining three fundamental constraints, the sample space was reduced from 1×5¹⁴ to 87 091 200, enabling tractable modeling. A Python-based algorithm was developed to randomly generate 4 500 core loading configurations, and validated SSF values for each were computed using the square fuel assembly core management system (CMS) code to establish a neural network dataset. Input features (core loading patterns and rod sequences) underwent preprocessing: 1) conversion into a 26×1 vector (loading configuration) and a 1×226 vector (rod sequence), merged into a time-sequential matrix; 2) normalization to eliminate scale-driven dominance; 3) dataset partitioning into training and test sets at an 8∶2 ratio. A convolutional neural network (CNN) model was constructed on the TensorFlow framework to map inputs to SSF outputs. Critical components including activation functions, optimizers and loss functions, were selected through empirical studies. Hyperparameters (convolutional layers, epochs, learning rates, batch sizes) were optimized by evaluating MSE (mean squared error) and MAE (mean absolute error) metrics, defining the final network architecture and training parameters for the CNN-based SSF predictor. Numerical results show that training and testing MSE values decline sharply within 5 epochs, stabilizing near 5.70×10⁻⁵, while MRE (mean relative error) stabilizes after 20 epochs, with training MRE primarily ＜4% and testing MRE concentrated within 0%-2%. Over the final 20 test epochs, the maximum MRE remains below 2.25%, with an average MRE of 1.20%. For the actual Cycle 2 core configuration, all predicted SSF values (226×9 data points across 9 rod groups) show absolute relative deviations ＜2% against CMS benchmarks. The constructed CNN model demonstrates high prediction accuracy, stability, and generalization capability, indicating the feasibility of applying the CNN method for SSF computation in engineering practice.