Research On Boiling Heat Transfer Of Downward Facing Surface Based On Machine Learning

When a serious accident occurs in a pressurized water reactor nuclear power plant,if the heat of core melt cannot be effectively exported,the lower head of a reactor pressure vessel(RPV)will be melted through,thereby threatening the integrity of containment.As the last barrier to prevent radioactive leakage,once the containment’s integrity is damaged,serious radioactive accidents will occur.In Vessel Retention(IVR)strategy is used by advanced pressurized water reactors,including HPR1000,CAP1400,and AP1 000,to retain core debris inside pressure vessel.External Reactor Vessel Cooling(ERVC),which removes heat from core debris by injecting water into the outer cavity of the pressure vessel,is an important method for implementing IVR strategy.The key to the success of IVR-ERVC is to ensure that the heat load at different positions of the lower head is lower than the critical heat flux(CHF)of boiling heat transfer at the corresponding position.Therefore,the boiling heat transfer on downward facing surface has become one of the hot spots in the field of nuclear power plant safety technology at home and abroad.After decades of development,a considerable amount of experimental data has been accumulated on downward facing boiling heat transfer,but there are still many difficulties,for example,the application range of the CHF correlation formula proposed so far is limited,traditional methods require manual identification of boiling heat transfer zones and are difficult to achieve real-time heat flux prediction through non-contact measurements.In recent years,with the increase of computer processing capabilities,the potential application of machine learning methods in the field of boiling heat transfer has been discovered by researchers.Therefore,based on the machine learning method,this study will obtain a prediction model of the pool boiling CHF on downward facing surface with a wider prediction range.This study will also record the boiling process through experiments,and use the machine learning method to analyze the dynamic image of bubbles,so as to realize the classification of boiling states and the prediction of heat flux.In this study,we compared the accuracy of random forest,ε-support vector machine(ε-SVM)and multilayer perception(MLP)in predicting CHF based on a large amount of CHF data collected in published papers on heat transfer under pool boiling,the results showed that ε-SVM model could provide the lowest mean absolute percentage error(MAPE)of 2.0785%for test data,and could reveal the impact of different parameters on CHF,moreover,the prediction accuracy of this model was higher than the CHF correlation proposed by current researchers,and its application range was also wider.This study provides a new idea for the prediction of pool boiling CHF on the downward facing surface.This study also compared the ability of MLP and convolutional neural network(CNN)to classify pool boiling states.The results showed that CNN could classify images with higher accuracy at lower image resolution,even if the image resolution was as low as 16 px × 9 px,as long as the parameters were set reasonably,CNN could still complete classification with an accuracy of 99.7%.The number of neurons in the fully connected layer,the number of convolution kernel and the image resolution had coupling effects on the model.When the image resolution was 16 px × 9 px,the prediction accuracy of the model increased with the increase of the inclination angle,which was due to the fact that the closer the boiling surface approaches the horizontal downward direction,the easier it was for the bubbles to gather into clusters and exhibit a flat feature,making it difficult to escape from the surface,and at this time the image contains fewer features.In addition,the MLP model could predict heat flux based on images containing bubble information.For test images with inclinations of 30°,45°,and 60°,the predicted MAPEs are 4.51%,3.13%,and 4.89%,respectively.In this study,the heat flux was also predicted by sections according to the heat flux value,the results showed that the accuracy of the model was higher when segmented prediction was performed,and the prediction deviation in the high heat flux section was higher than that in the low heat flux section.The existing research results were of great value for in-depth understanding of the effects of various parameters on pool boiling heat transfer CHF on downward facing surface,and provided an important basis for machine learning methods in the field of downward surface boiling heat transfer,especially for image based boiling heat transfer region classification and heat flux density prediction.