Study On Flow And Heat Transfer Characteristics Of Supercritical CO₂ Based On Neural Network

Supercritical Carbon dioxide（S-CO₂）is widely used in the GenⅣnuclear power and solar thermal power plants,as well as the cooling of high heat flux surfaces,because of its special thermophysical properties.The Printed Circuit Heat Exchanger（PCHE）in S-CO₂Brayton cycles and S-CO₂ jet impingement cooling on high heat flux surface are investigated in this study.The present thesis studies the flow and heat transfer characteristics of S-CO₂ based on neural network in the context of the existing heat transfer correlation could not accurately calculate local heat flux.Firstly,the CFD data and BP neural network are combined to establisha method for the fast prediction of local flow and heat transfer characteristics of S-CO₂ in sinusoidal channels.The non-dominated sorting genetic algorithm（NSGA-II）with elite strategy is used to optimize the heat transfer rate and pressure drop of the heat exchanger.The optimization direction for the design of the sinusoidal channel PCHE is also obtained based on cost analysises.Secondly,the BP neural network is further used under the background of S-CO₂ jet impingement cooling.A method nesting two neural networks is proposed to predict the heat flux fields of S-CO₂ multiple impinging jets.The main research contents and conclusions of this thesis are as follows:（1）A fast prediction method to predict local Nu and f factor in sinusoidal channels is established.After screening the input parameters of neural network,the neural network model structure is determined.Then,the sub-heat exchanger model is used to predict the heat transfer rate Q and pressure dropΔP in different channel structures.The results show that the established prediction method has high accuracy in predicting local Nu and f factors in different sinusoidal channel structures and the corresponding R² could reach 0.952 and 0.982,respectively.Compared with the calculation results of heat transfer rate Q and pressure dropΔP based on the commonly used global correlations,the accuracy of the prediction method in this thesis can be improved by 2%～10%.（2）By combining the fast prediction method with NSGA-II,the optimal sinusoidal channel structure parameters are calculated under the constraints of maximum heat transfer rate and minimum pressure drop.Then,the optimization direction for sinusoidal channel PCHE design is obtained through the following cost analysis based on the optimization results.The resulted Pareto optimal solution reveals that the heat transfer rate of the PCHE is more sensitive to the change of channel length and the pressure drop is more sensitive to the change of angle.The cost analysis with fixed payback period of 30 years shows that the total cost rises faster with the increase of angle.However,under variable payback period conditions,the annual total cost increases at a slower rate with the year increases.（3）A fast prediction method for the prediction of heat flux fields of S-CO₂ multiple impinging jets is constructed by nesting two neural networks which has improved the prediction accuracy of the heat flux fields.Then the thesis uses test cases in different research scopes to verify the accuracy of the fast prediction method.The results show that,compared with the prediction results of the neural network model which uses the variables of height,mass flow rate and heat transfer surface coordinate as input parameters,the prediction accuracy of fast prediction model is greatly improved.The first neural network is obtained by training single nozzle jet data and the R² is above 0.9 while the R² of the second neural network trained with multi-nozzle is more than 0.6（some cases can reach more than 0.85）.The accuracy of validation cases within the variable research scope is higher than that of outside the variable research scope.（4）By adjusting the normalization range of input parameters,the accuracy of the validation case which below the data range of the training set is improved.The accuracy of the fast prediction results by using two neural network nesting models is verified.Then the results compared with other machine learning prediction results.The results show that,changing the traditional normalization range so that all normalized variables used for training and verification cases become positive values,then compared with the previous results,the R² of single nozzle and multi-nozzle model improve above 0.2.The edge of high heat flux become more smooth in contours which is consistent with the actual model.Compared with other models,the prediction model used in this paper has higher accuracy.