A long-term dependable and reliable method for reactor accident prognosis using temporal fusion transformer

Abstract

Introduction: The accurate prognosis of reactor accidents is essential for deploying effective strategies that prevent radioactive releases. However, research in the nuclear sector is limited. This paper introduces a novel Temporal Fusion Transformer (TFT) model-based method for accident prognosis that incorporates multi-headed self-attention and gating mechanisms.Methods: Our proposed method combines multi-headed self-attention and gating mechanisms of TFT with multiple covariates to enhance prediction accuracy. Additionally, we employ quantile regression for uncertainty assessment. We apply this method to the HPR1000 reactor to predict outcomes following loss of coolant accidents (LOCAs).Results: The experimental results reveal that our proposed method outperforms existing deep learning-based prediction models in both prediction accuracy and confidence intervals. We also demonstrate increased robustness through interference experiments with varying signal-to-noise ratios and ablation studies on static covariates.Discussion: Our method contributes to the development of intelligent and reduced-staff maintenance methods for reactor systems, showcasing its ability to effectively extract and utilize features of static and historical covariates for improved predictive performance.