A Transformer-VAE Framework with Knowledge Distillation for Fast Prediction of Nuclear Power Plant Accident Transient Response

Conventional analysis of nuclear power plant accident transient responses heavily relies on physical simulation programs, whose computational time significantly exceeds the actual accident response duration, thereby severely hindering real-time safety assessment. This paper proposes a novel framework that integrates a Transformer-based Variational Autoencoder (VAE) with knowledge distillation for the fast prediction of nuclear power plant (NPP) accident transient responses. The approach involves constructing a latent space to extract essential features from transient response data using an Encoder–Decoder model based on the VAE architecture. A key innovation is the establishment of a direct mapping between plant operating condition parameters and the latent space, enabling the one-step generation of accident transients without iterative sequential prediction. The Encoder and Decoder leverage Self-Attention and Cross-Attention mechanisms to enhance feature extraction and conditional generation. Furthermore, the Encoder is distilled into a Mapper network, which predicts the latent features directly from the operating conditions, resulting in an efficient Mapper–Decoder pipeline for rapid prediction. The proposed model was evaluated against traditional Long Short-Term Memory (LSTM) and fully connected neural networks. Experimental results demonstrate that the proposed framework achieves superior performance in predicting NPP accident transients, indicating its strong potential for efficient safety analysis and system design optimization.