Hybrid neural network and statistical forecasting methodology for predictive monitoring and residual useful life estimation in nuclear power plant components

Abstract

Lifetime extension of key components of Nuclear Power Plants (NPPs) is of great importance for reliable and continuous energy production. In this regard, the present paper proposes a novel Neural Network (NN) and time-series forecasting-based approach for the prediction of the health condition of nuclear components and the estimation of Remaining Useful Life (RUL) regarding Class II components, focusing on critical piping systems that may be subject to Flow- Accelerated Corrosion (FAC). A digital replica of a Class II piping was implemented in a finite element code to simulate the progressive thinning it suffers because of operational and environmental conditions. To the aim of the present study, generated synthetic data have been employed in the training phase of the NN model. Autoencoder, which is a special type of NN, that compresses input data into one major holistic signal representing the component’s health status, is used. The component RUL is computed in this study by the ARIMA algorithm. Results showed that the adopted hybrid methodology is capable of forecasting accurately the piping plastic deformation 6 to 14 months in advance, thus allowing for better and more efficient NPP maintenance and management.