A robust approach for time-series modeling and forecasting in nuclear power plants through an improved transformer model and Bayesian optimization

As a highly reliable base-load energy source, nuclear power plants (NPPs) have served as a critical platform for the advancement of prognostics and health management (PHM) technologies. The safe operation of NPPs critically relies on the accurate forecasting of key operational parameters through time-series analysis. While neural network-based models have demonstrated substantial efficacy in predictive tasks, they often encounter limitations when modeling the complex and highly nonlinear dynamics of NPP systems, particularly in capturing long-term dependencies. This study addresses the multivariate time-series forecasting challenge by implementing the INV-Transformer model on a dataset simulating a loss of AC power scenario in NPPs. The proposed model architecture, enhanced with Bayesian optimization, significantly improves prediction accuracy and model adaptability. Experimental results indicate that the INV-Transformer model effectively predicts critical physical parameters, reducing the mean absolute error from 0.0394 to 0.0146 (a reduction of 62.9 percent) and the mean squared error from 0.0028 to 0.0003 (a reduction of 89.3 percent). These findings establish the INV-Transformer as a novel and robust approach for predictive modeling in the operational management of NPPs.