基于深度学习的超临界二氧化碳临界流模型研究
Research on Supercritical Carbon Dioxide Critical Flow Model Based on Deep Learning

破口处的临界流量决定了冷却剂丧失速度和一回路泄压速度，二氧化碳(CO₂)在拟临界附近会发生物性畸变，而超临界二氧化碳(SCO₂)反应堆的喷放过程涉及跨临界降压、不同相态喷放流量计算，现有研究缺乏宽参数范围高精度临界流模型。数据驱动方法可基于成长型的训练数据库提升精度，为了快速且精准地预测SCO₂临界流，文章基于深度学习方法建立了SCO₂临界流模型。通过敏感性分析确定了深度学习模型的特征输入；以循环神经网络(RNN)为框架，使用K折交叉验证、L2正则化得到了预测精度更高、泛化能力更强的临界流模型SCO₂-RNN；基于遗传算法选取最优超参数，得到模型对预测结果平均相对误差为4.88%，最大相对误差为14.24%；对初见数据泛化的平均误差为5.73%，最大误差为20.45%；在使用迁移学习后，平均相对误差降低为1.75%，最大相对相对误差为4.15%。基于训练好的模型对新数据的泛化结果表明：深度学习模型在速度和精度方面均满足工程要求，可建立宽适用范围的高精度高效率临界流模型。
The critical flow at the break determines the speed of coolant loss and the depressurization rate of the primary circuit. Carbon dioxide (CO₂) reactor undergoes physical property anomalies near the pseudo-critical point, and the supercritical carbon dioxide (SCO₂) discharge process involves cross-critical pressure reduction and calculation of discharge flow rates in different phases. Existing research lacks a high-precision critical flow model with a wide range of parameters. Data-driven methods can improve accuracy based on a growing training database. In order to quickly and accurately predict the critical flow of supercritical carbon dioxide, this paper establishes an SCO₂ critical flow model based on deep learning methods. Through sensitivity analysis, the feature input of the deep learning model was determined; using the Recurrent Neural Network (RNN) as the framework, with K-fold cross-validation and L2 regularization, a critical flow model SCO₂-RNN with higher prediction accuracy and stronger generalization ability was obtained. Then we selected the best hyperparameters were selected by the genetic algorithm. The average error of the predicted result of this model is 4.88%, and the maximum error is 14.24%; the average error of generalizing to unseen data is 5.73%, and the maximum error is 20.45%. After using transfer learning, the average error decreased to 1.75%, and the maximum error is 4.15%. This indicates that with the upgrade of database, there will be higher accuracy and better adaptability. The generalization results of new data based on the trained model show that: the deep learning model meets engineering requirements in terms of speed and accuracy, and can establish a high-precision and high-efficiency critical flow model with a wide range of applicability.