The Study Of Fuel Property Prediction And Sensitivity Analysis Based On Artificial Neural Network

Environmental pollution and fossil fuel reserves are the two main problems of traditional energy nowadays.The solution,on the one hand,is to reduce the pollution emissions from traditional energy sources.And on the other hand,is to develop green and environmentally friendly new energy sources.In the current situation where traditional energy accounts for the bulk of energy consumption and new energy requires time buffering,it is necessary to improve the performance indicators of traditional fuels,such as increasingly strict emission indicators.The various indicators of fuel need to be measured through standard operating conditions experiments.For example,in the production process of refined oil,various physical properties need to be monitored during various processes.Traditional experiments have disadvantages such as large sample consumption,time-consuming,and poor error reproducibility.There is an urgent need for chemometric methods for fast and efficient measurement calculations.The purpose of this paper is to construct a model for predicting fuel physical properties based on measured component data and use sensitivity algorithms to explain the correlation between components and physical properties.This paper uses neural network algorithms to build a predictive fuel physical property model.The neural network model is a kind of empirical model.The advantage is that it can automatically adjust the model to adapt to the problem through learning.As long as there is enough data,it can fit the problem well.The disadvantage is that it requires a lot of data.This process is similar to the numerical simulation process of physical equation modeling.The numerical simulation requires the law of the research problem to be known.Through a series of control equations,the behavior of the research object in the system is modeled coherently.The advantage is that the physical meaning of the modeling process is clear.The disadvantage is that you need to know the governing equations.If you model according to the first principles,it will consume a lot of time.The empirical model is very suitable for fitting nonlinear complex relationships.In terms of sensitivity analysis,an algorithm based on solving partial derivatives is selected.This method is suitable for situations where the input data distribution is unknown and the model is easy to be derived.For the analysis of the component data of diesel,the distillation range and hydrocarbon composition were selected as input features,and the prediction models were constructed respectively,and the Mat Lab GUI interface program was written to apply the models in practice.The relative error of the model testing set is between 0.2%and 2.5%,which meets industrial-grade requirements.For biodiesel,the component input and functional group input are selected according to component characteristics.The FAME components affect physical properties by their molecular characteristics,proportions,and coupling relationships between molecules.The input of functional groups also has such an influence on physical properties.At the same time,the functional group determines the influence of the molecule itself on physical properties.For component input and functional group input,the average absolute errors of the training set and testing set are 1.04 and 0.91,respectively,and the average relative errors are 1.88%and 1.60%,respectively.Through sensitivity analysis,it is concluded that C18:02,C18:01,and C16:00have the strongest single-factor sensitivity to CN.The sensitivity of n(-CH₂-)/n(C)is greater than n(C=C)/n(C).C18:02 and C18:01 have high sensitivity due to the large amount of C=C.In the two-factor sensitivity,since C=C and-CH₂-have opposite effects on CN,when combined with C18:01 and C18:02,respectively,the two-factor sensitivity coefficients of C16:00 and C18:00 are greater than C18:03;When coupled with position index of C=C,the sensitivity of n(-CH₂-)/n(C)is greater than n(C=C)/n(C).