Neutronics Optimaization Of Fuel Assembly Based On Data Mining

The performance optimization of nuclear fuel assembly is a typical multivariable and multi-objective optimization problem.Two problems would be solved that the design of fuel assembly mainly faces.one of them is that the time of high-precision multi-group transport burnup calculation for the assembly is long,and the other one is only a small number of assembly core schemes be selected and designed manually.In view of the above two problems,this paper took the common PWR fuel assemblies as the research object,and carries out the prediction of the neutron physical parameters of the fuel assemblies and the core scheme selection based on the assembly parameters.Firstly,the key factors of the neutron performance of fuel assembly are analyzed,and the optimization design of fuel assembly based on data mining technology was explored.Then it is determined to use Linear Regression,Support Vector Machine?LibSVM?,Decision Tree,Multi-layer Perceptron?MLP?and Random Forest algorithm to carry out parameter prediction and scheme screening research.Secondly,aiming at the long calculation time of high-precision assembly transportation and burn-up,linear regression,decision tree,multi-layer perceptron and random forest algorithm were used to predict the average heat flux and infinite multiplication coefficient k_inf of fuel assemblies at the beginning and end of their life cycle,and their accuracy was verified.The results show that the prediction time of the four kinds of algorithms was within 3S;compared with the other three algorithms,the prediction ability and convergence of the model of random forest is the best,and the error was acceptable,followed by the decision tree model,which verifies the feasibility of using data mining technology to predict the neutron physical parameters.At the same time,the sensitivity of the fuel concentration of the module to the above four target parameters is the highest than burnable poisons arrangement and content,and the sensitivity of the arrangement and content of burnable poisons is different according to the different life span.Then,aiming at the problem that the core pre-set scheme is less and mainly relies on artificial experience to search and select,the decision tree,support vector machine,multi-layer perceptron and random forest algorithm are used to carry out the assemblies-core scheme selection research with the fuel enrichment in different areas and the type of burnable poisons as the characteristic variables to select.The objective function was selected as Uneven coefficient of keff during lifetime?UCKL?,Radial power non-uniformity coefficient?RPNC?,Radial flux non-uniformity coefficient?RFNC?and the core life?CL?,and the objective function compliance CPF is taken as the overall planning.The research results show that the highest training accuracy is random forest model,the lowest is LibSVM,the longest time to train the model and predict is MLP,and the shortest is C4.5 model;in the case classification of CPF=4,there are problems of high accuracy,low completion or high completion but low accuracy;then through the simplification and adjustment of logarithmic data model,the four algorithms could complete the selection of component schemes well,and the accuracy of prediction is over 94%and the time is less than 10s,which proves the feasibility of using data mining technology to carry out rapid screening of component core scheme.At the same time,all assemblies types of the entire core are divided into three zones A,B and C from inside to outside.Through sensitivity analysis,the sensitivity of fuel assembly in Zone C is higher than that in Zone A and B;that is to say,Zone C contributes more to the overall nonuniformity of the core.In addition,in the component design variables for zone A,the most sensitive to CPF results is the fuel concentration,while the highest in Zone C is the arrangement of burnable poisons.