Time series modeling and forecasting with feature decomposition and interaction for prognostics and health management in nuclear power plant

Advanced nuclear power systems have raised higher demands for safety, driving the rapid development of prognostics and health management (PHM). However, despite significant progress in time series forecasting technology for PHM in nuclear power plant (NPP), challenges remain in complex system modeling and long-term dependency modeling. To address these issues, this paper proposes an end-to-end multivariate multi-step ahead time series forecasting method, named Attention-guided Feature Decomposition and Interaction Network (AFDI-Net). Firstly, a multi-kernel time series decomposition module with attention mechanism is used to decompose operating parameters of NPP into multiple trend and seasonal patterns, and autonomously learn the weight of each pattern. Secondly, a feature reconstruction module with information interaction is introduced to model the local features and global correlations, enhancing feature representation and long-term modeling capabilities. Finally, a simple linear layer is used for time series forecasting, mapping the features to the required prediction window. In the experimental phase, given the limitations of traditional evaluation metrics that only assess methods from a numerical perspective, an evaluation system based on downstream task is introduced. Extensive experimental results based on the digital Instrumentation and Control system show that compared with state-of-the-art methods, AFDI-Net yields 41.5% and 20.7% relative improvements for short-term and long-term time series forecasting respectively, and exhibits good robustness.