Research On Prediction Model Of Multi-component Liquid Rollover Based On Neural Network

Based on the liquefied natural gas(LNG)rollover problem as the engineering background,this paper studies the prediction model of the critical conditions,rollover time and duration of multi-component liquid rollover.Rollover is a kind of accident that is easy to happen in the process of LNG storage and transportation.When there is density stratification in LNG storage tank,methane in the upper LNG tank is preferred to evaporate under the action of environmental heat leakage,and the content of liquid recombination increases relatively and the density increases.As the layered interface blocks the natural convection movement between liquid layers,only a small amount of heat absorbed by LNG from the lower layer is transferred to the liquid from the upper layer through thermal conduction,resulting in the rising temperature and decreasing density of LNG from the lower layer.When the boundary between the liquid layers is unstable,LNG will roll and improper control will pose a threat to the safety of LNG storage and transportation as well as the surrounding environment.Therefore,in order to prevent the occurrence of LNG roll,the prediction and early warning of LNG roll should be strengthened.This paper USES the genetic neural network to study the roll prediction model,and the main work and conclusions are as follows:(1)Build a neural network model for predicting the occurrence conditions,occurrence and duration of multi-component liquid tumbling,optimize the initial weights and thresholds of the neural network with genetic algorithm,and write the genetic neural network program code with MATLAB.(2)with layered multicomponent liquid rollover process based on computational fluid dynamics simulation example,through the analysis of sidewall boundary lifting flow velocity,the changing rule of the mainstream zone flow field and concentration distribution and migration characteristics of layered interface,establishes the initial buoyancy ratio,wall heat leakage,initial concentration,initial density difference as input index,and tumbling on poor critical concentration and the critical density difference occurred as output index system;By extracting the characteristic indexes of rollover time under different initial conditions,28 groups of sample data were obtained and normalized.The samples were input into the neural network for training and optimization.The results showed that the best training effect was achieved when the population size was 400,the evolutionary algebra was 400,and the number of hidden layers was 9,forming a 4-9-3 neural network structure.Through the test sample analysis,the simulation error of weight/threshold after using random weight/threshold and genetic algorithm optimization shows that the optimized network has better fitting and the prediction error is reduced by 83.63%.The maximum relative error of critical density difference compared with the original value was 7.59%,the maximum relative error of critical concentration difference was 7.29%,and the maximum relative error of rollover time was 12.17% when the two groups of LNG simulation results were substituted into the neural network.(3)based on the analysis of the rollover process,the concentration distribution in cloud boundary penetrating speed and the concentration of the central tank and density change law,set up to wall heat leakage rate,poor concentration,critical density difference,roller critical penetration rate,maximum penetration rate as input index,with rollover duration as output indicators index system;By extracting the characteristic indexes of the rollover process under different initial conditions,28 groups of sample data were obtained and normalized.The samples were input into the neural network for training and optimization.The results showed that the training effect was better when the population size was 400,the evolutionary algebra was 400,and the number of hidden layers was 9,and the optimal network structure was 5-9-1.By using the random weight/threshold and the genetic algorithm to optimize the weight/threshold simulation error through the test sample analysis,it shows that the optimized network accuracy is higher and the prediction error is reduced by 18.37%.When the calculated results of the two groups of LNG were input,the maximum relative error of roll duration was 5.54%.