A reliable and adaptive prediction framework for nuclear power plant system through an improved Transformer model and Bayesian uncertainty analysis

With the increasing adoption of digital transformation in nuclear energy, digital twin technology for nuclear power plants has emerged as a powerful tool for system management. A key function within this framework is accurate time-series prediction. However, current models in the nuclear field, primarily based on neural networks, often require extensive training data and face limitations in multi-parameter and multi-step predictions. Furthermore, dependence on simulated data can result in performance gaps when applied to real-world scenarios. To address these challenges, this paper proposes a multi-parameter and multi-step pre-trained Transformer model, enhanced through transfer learning and incremental learning strategies, to improve prediction accuracy and adaptability. In addition, a robust uncertainty analysis framework is integrated to quantify and manage the uncertainties inherent in model predictions. This integration of digital twin technology with advanced time-series prediction models provides a novel and reliable approach to improving nuclear power plants’ operational safety and efficiency.