【摘要】 虚拟测量方法常被用于核反应堆瞬态工况监测.基于数据驱动方法，虚拟测量方法不直接依赖传感器获取的数据，能够解决传统监测方法部署成本高、维护困难等问题.当前，主流虚拟测量方法往往存在特征捕获能力不强、预测精度不足等问题.本文构建了一种融合注意力机制与ED-LSTM(Encoder-Decoder LSTM)模型的虚拟量测方法 .基于PCTRAN仿真软件生成的高保真核反应堆动态数据集，本文分别将时间注意力、因果自注意力、卷积注意力及分层注意力等4种注意力机制引入ED-LSTM模型，以增强ED-LSTM模型对关键时序特征的提取能力.其中，引入注意力机制的方式有3种，即只在编码器添加、只在解码器添加以及同时在编码器和解码器添加.为获得最佳模型参数值，本文设计了13种方案，分别进行仿真，并通过均方根误差（RMSE）、平均绝对误差（MAE）和判定系数（R2）等指标对模型的预测性能进行评价.结果显示：（i）在编码器中添加各种注意力机制都能提升模型的预测性能，其中添加融合时间注意力机制的效果最好（RMSE降低23.4%）;(ii)以不同方式添加因果注意力机制后，模型的预测性能均有提升且效果较稳定；（iii）在解码器中添加时间、卷积或分层注意力机制导致模型的预测性能下降，可能原因是存在信息冗余或过拟合问题.本文的研究表明，将注意力机制引入ED-LSTM模型、提升虚拟测量方法的精度是可行的.更多还原

【Abstract】 Virtual measurement(VM) approaches are frequently employed in nuclear engineering for the transient condition monitoring of nuclear reactor. As a data-driven approach, VM eliminates the reliance on physical sensors and effectively overcomes the limitations of conventional monitoring techniques, such as high deployment costs and maintenance challenges. Nowadays, mainstream VM approaches still exhibit inadequate temporal feature extraction and suboptimal prediction accuracy. In this paper, an enhanced VM framework that integrates attention mechanism with Encoder-Decoder Long Short-Term Memory(ED-LSTM) architecture is proposed. Four attention mechanisms are used: temporal attention, causal attention, convolutional attention and hierarchical attention. Meanwhile, the attention mechanisms are incorporated into the EDLSTM model in three ways: encoder-only, decoder-only and encoder-decoder hybrid. To optimize the model parameters, a high-fidelity nuclear reactor transient dataset generated by PCTRAN simulation software is used, and 13 parameter configuration schemes are evaluated by using the performance metrics including Root Mean Square Error(RMSE), Mean Absolute Error(MAE) and coefficient of determination(R2), respectively. Simulation results demonstrate that:(i) The integration of every attention mechanism into the encoder of ED-LSTM model enhances the model prediction performance, in which the temporal attention mechanism achieves the optimal result through a 23. 4% RMSE reduction;(ii) The integration of causal attention mechanism in every way improves the prediction stability across implementations;(iii) The integration of temporal, convolutional or hierarchical attention mechanism into the decoder of ED-LSTM model degrades the model prediction performance, likely due to the information redundancy or overfitting phenomena. The obtained results substantiate the technical feasibility of integrating attention mechanism with ED-LSTM architecture for the precision enhancement of VM approaches.更多还原