Among the current mainstream nuclear reactor thermal-hydraulic calculation system analysis softwares (such as RE-LAP5, TRACE, etc.), different empirical or semi-empirical relationships are employed to calculate two-phase interfacial flow parameters. However, due to the lack of understanding of the physical laws of reactor fluids and the limited experimental data currently, the application scope and accuracy of these methods are relatively limited. As an important paradigm in machine learning methods, ensemble learning has received great attention in numerical prediction tasks in many fields due to its excellent robustness, nonlinear fitting ability and strong interpretability. In this study, three typical ensemble learning methods including random forest, gradient boosting regression tree, and extreme gradient boosting tree are used to build a two-phase interfacial parameter prediction model for rectangular channels. The investigated interfacial parameters include interfacial area concentration, void fraction, bubble velocity, bubble chord length, and bubble frequency. The result reveals that XGBoosting has the best performance in forecasting. In order to further improve the performance of the ensemble learning method and solve the problem of poor generalization ability of the homogeneous-ensemble model, the heterogeneous-ensemble strategy are adopted to combine different types of algorithms to provide a model with stronger scalability and generalization ability for two-phase flow interfacial parameter prediction. The results show that the usage of the heterogeneous-ensemble method can improve the prediction ability of the model to a certain extent, and further ensure the reliability of the data-driven model for the prediction of the two-phase interfacial flow key parameters in the rectangular channel.

You do not currently have access to this content.