The accuracy of transient thermal-hydraulic parameters within different operational states of a reactor core is crucial for reactor safety. Rapid and precise prediction of key thermal parameter trends is essential for enhancing reactor safety.

This study aims to propose a novel prediction method of reactor transient thermal-hydraulic parameters based on Sequence-to-Sequence (Seq2Seq) neural network model for improving the accuracy and speed of predicting thermal parameters to ensure the safe operation of nuclear power plants.

Firstly, Long Short Term Memory (LSTM) neural network was coupled with the Convolutional Neural Networks (CNN) to form a Seq2Seq (Sequence to Sequence) neural network model, and the wavelet decomposition was applied to preprocessing thermal parameter data. Then, the SUBCHANFLOW sub-channel program was employed to generate data samples from the China experimental fast reactor (CEFR), and results were comprehensively evaluated using the rank-sum ratio (RSR) method to derive an optimal prediction scheme. Finally, the generalization ability of this scheme was further assessed through time-series-based K-fold cross-validation and bootstrapping methods.

The coupled CNN-LSTM Seq2Seq neural network model exhibits superior predictive performance, with high accuracy and robust fitting capabilities. The maximum average relative error recorded is 0.552%.

The developed Seq2Seq model in this study efficiently extracts time series features and demonstrates strong generalization capabilities, providing a valuable reference for predicting critical thermal parameters in reactors.