Yanjie Tuo Xiaojing Liu ^*

• Shanghai Jiao Tong University, Shanghai, China

In the domain of nuclear power plant operations, accurately and rapidly predicting future states is crucial for ensuring safety and efficiency. Data-driven methods are becoming increasingly important for nuclear power plant parameter forecasting. While Transformer neural networks have emerged as powerful tools due to their self-attention mechanisms and ability to capture long-range dependencies, their application in the nuclear energy field remains limited and their capabilities largely untested. Additionally, Transformer models are highly sensitive to data complexity, presenting challenges for model development and computational efficiency. This study proposes a feature selection method that integrates clustering and mutual information techniques to reduce the dimensionality of training data before applying Transformer models. By identifying key physical quantities from large datasets, we refine the data used for training a Transformer model, which is then optimized using the Treestructured Parzen Estimator algorithm. Applying this method to a dataset for predicting a shutdown condition of a nuclear power plant, we demonstrate the effectiveness of the proposed "feature selection + Transformer" approach: (1) The Transformer model achieved high accuracy in predicting nuclear power plant parameters, with key physical quantities such as temperature, pressure, and water level attaining a normalized root mean squared error below 0.009 , indicating that the RMSE is below 0.9% of the range of the original data, reflecting a very small prediction error. (2) The feature selection method effectively reduced input data dimensionality with minimal impact on model accuracy.This paper is structured as follows: Section 2 introduces the components of the proposed method, including the framework, feature extraction methods, neural network model, and optimization algorithm. Section 3 presents an overview of the operational data from NPPs used in this study. Section 4 discusses the results obtained from this method. Finally, Section 5 presents the conclusions drawn from the study.