Research On Algorithm And Application Of Convective Heat Transfer Problem Based On Deep Learning

Thermal-hydraulic processes are widely present in nuclear power engineering systems,and the heat transfer characteristics of gas-liquid two-phase flow in narrow slit channels in nuclear power plants are very complex,and the acquisition of its information can help us better analyze the thermal-hydraulic processes occurring in the reactor and timely assess and control the accident risks.However,due to the small size of narrow-slit pipes and very complex gas-liquid structure,the problem of convective heat transfer inversion of narrow-slit channels needs to be solved by traditional means of parameter measurement,which has a single measurement object,cannot be predicted in real time,and is complicated to operate and difficult to guarantee the accuracy.In recent years,deep learning techniques have made significant progress in tasks such as bubble dynamics information acquisition and flow field reconstruction,but there are still problems such as strong dependence on image data,lack of model interpretability and unstable model training.To address the above shortcomings,this paper combines deep learning algorithms with traditional experimental methods and classical numerical simulations,and applies deep learning techniques based on different principles to convective heat transfer problems to solve the problem of rapid acquisition of physical field information in narrow slit blind areas.The main contents and research results of this paper are as follows:(1)To address the problem of difficulty in obtaining bubble information in pipes,a convolutional self-encoder model based on unsupervised learning is constructed to extract features from experimental data of different gas-liquid two-phase flow patterns,and three different machine learning classification algorithms are used for flow pattern recognition,which solves the problems of difficult feature extraction,incomplete extracted information and low accuracy in flow pattern recognition methods,reduces the classification model training process on a large number of It reduces the dependence of the classification model training process on a large amount of label data,and improves the accuracy of the machine learning algorithm stream-type recognition and the generalization ability of the model.(2)To further obtain more information about the flow field inside the pipe of gasliquid two-phase flow,a conditional generation adversarial neural network model based on image recognition is constructed to predict the flow state of the fluid inside the pipe,achieving fast and accurate prediction from the geometric profile of bubbles inside the pipe to the flow field and pressure field inside the pipe.In terms of computational accuracy,the relative error of prediction on the test set is less than 10%;while in terms of computational efficiency,the prediction speed is 10 times faster than that of numerical simulation and simulation,solving the problem that fluid flow state information in narrow slit channels is difficult to predict accurately and quickly.(3)To address the problems of high cost of image data acquisition and single prediction parameters of deep learning algorithm,the physical information neural network model is constructed,and the physical information neural network is applied to achieve accurate and fast prediction of multiple coupled physical fields such as velocity field and pressure field in gas-liquid two-phase pipeline using only boundary conditions and control equations.It provides a solution with high accuracy and lower training cost to solve the problem of blind area measurement in narrow slit channels,solves the problem of difficult data acquisition for image recognition-based algorithms,and improves the prediction efficiency.(4)To solve the gradient pathology problem of physical information neural network in solving the convective heat transfer problem,the training process of the traditional physical information neural network method is improved,and the reliability of the model is verified.Numerical results show that the method combining adaptive weighting coefficients and learning rate annealing algorithm can better alleviate the problems of imbalance and gradient disappearance during the backward propagation of gradients,and effectively improve the accuracy of the physical information neural network in predicting convective heat transfer problems.