2026年“核理安邦”联合博士生学术论坛

Nuclear Accident Evacuation Path Planning Based on Transformer fused with physical information

Not scheduled

1h

地址：清华大学校内

海报展示 海报展示 海报展示

Speaker

马 悦翔 (中国辐射防护研究院)

Abstract

The dynamic radioactive plumes induced by nuclear accidents introduce significant uncertainties into off-site emergency evacuations. To address the vulnerability to radiation exposure and the lack of global spatiotemporal perception inherent in traditional algorithms,
this paper proposes a spatiotemporal path planning model for nuclear emergencies, termed Phy-TransA, which integrates physical priors with deep learning. The model employs a hierarchical decision-making architecture: it generates a set of candidate paths based on the A algorithm and road network topology, and utilizes a Transformer network to analyze multidimensional spatiotemporal physical feature vectors, predicting the risk probabilities of decision node branches across different spatiotemporal contexts. Radiation field simulations were conducted using the URBAN 2000 IOP #10 field tracer experiment data and a Lagrangian particle tracking model. The results indicate that in unknown heterogeneous
scenarios characterized by source spatial offsets and intermittent releases, Phy-TransA exhibits algorithmic validity and scenario adaptability. Compared with the traditional A algorithm, the proposed model reduces the median cumulative dose of evacuation paths by 95.40%, accompanied by a 38.28% increase in travel time, effectively bypassing local high-concentration plumes.

摘要

核事故引发的动态放射性烟羽给场外应急撤离带来了极大的不确定性。针对传统算法易导致高辐射暴露、以及全局时空感知能力不足的局限，本文提出了一种融合物理先验与深度学习的核应急时空路径规划模型（Phy-TransA）。该模型采用分层决策架构：基于A算法结合路网拓扑生成候选路径集，利用Transformer网络分析多维时空物理特征向量，预测不同时空下决策节点分支的风险概率。基于URBAN 2000 IOP #10现场示踪实验数据与拉格朗日粒子追踪模型的进行辐射场仿真，结果表明：在释放源存在空间偏移与间歇性释放的未知异构场景下，Phy-TransA展现出良好的泛化能力。相较于传统A算法，本模型在仅增加 38.28% 通行时间的条件下，将撤离路径的累积剂量中位数降低了 95.40%，有效规避了局部高浓度烟羽。