Session: 01-08: Nuclear Plant Operation, Modification, Life Extension, Maintenance and Life Cycle - VIII

Paper Number: 135879

135879 - Enhancing Nuclear Power Plant Operational Forecasting With Transformer Neural Networks: A Time-Series Data Approach 

Abstract: 

In the realm of nuclear power plant operation, precise prediction of operational parameters plays a crucial role in ensuring safety and efficiency. This study is dedicated to establishing a sophisticated link between various operational data from nuclear power plants and their real-time operating conditions, utilizing advanced Transformer Neural Networks for predictive analysis. The goal is to gain an in-depth understanding of future operational states and to trace the trajectory of data trends over time.

The first phase of this study emphasizes enhancing the prediction process's efficiency. Given the large volume of data generated during the operation of nuclear power plants, it is crucial to filter and select data types that have a significant correlation with operational states. To achieve this, this study employs a modified k-means clustering method, enriched with fusion coefficients. This method clusters operational conditions of the power plant based on time, generating time-series data that effectively capture the dynamic nature of plant operations.

Subsequently, this study utilizes mutual information methods, in tandem with the operational time-series, to identify key data series that exhibit a strong link to these conditions. These data series are then  preprocessed, with 60% allocated for the training set, 10% for the validation set, and 30% for the test set. These data segments are used to train a Transformer Neural Network, aiming to develop a robust predictive model for these specific data types.

A critical aspect of this study involves the selection of hyperparameters for the Transformer model. The Tree-structured Parzen Estimator (TPE) method is employed for this purpose, facilitating the creation of an effective Transformer model. This model showcases exceptional ability in predicting future operational states and data trends, based on historical data, thus enhancing the overall predictive accuracy in nuclear power plant operations.

One of the challenges this study addresses is the inverse relationship between the reduction in training set size and the predictive accuracy of the model in time-series data analysis. To overcome this, this study introduces a correction factor for the predictive errors of the Transformer model. This adaptation aims to boost the model's performance with varying training set sizes, thereby improving its generalizability and applicability in different operational scenarios.

In conclusion, while this study offers substantial insights into the field of predictive modeling in nuclear power plant operations, it also explores the potential of Transformer Neural Networks in analyzing and forecasting complex operational conditions. By leveraging the intricate patterns in time-series data, this study aspires to contribute towards the advancement of predictive accuracy in industrial environments, particularly where precision and reliability are crucial.

Presenting Author: Yanjie Tuo Shanghai Jiao Tong University

Presenting Author Biography: Yanjie Tuo is currently pursuing the Ph.D. degree with the School of Mechanical Engineering, Shanghai Jiao Tong University, China.

Authors:

Yanjie Tuo Shanghai Jiao Tong University
Xiaojing Liu Shanghai Jiao Tong University