Application Of Big Data Analysis In Fuel Property Prediction,Reaction Mechanism Construction And Auxiliary Engine Calibration

With the development of industrial intelligence,the amount of work data generated in situations such as energy production,fuel property measurement,and equipment operation in the combustion and energy industry is increasing explosively.Nowadays,this area has become an important field of data analysis and big data technology application.In this paper,the active subspace method,artificial neural network and other big data analysis methods were applied in three aspects of researches:fuel property prediction,reaction mechanism construction,and auxiliary engine calibration.In the aspect of fuel properties prediction,1020 sets of experimental data and 902 molecular topological indexes were used to build the prediction model.The standard enthalpies of formation of hydrocarbons were well predicted by the active subspace method with the R~2 being 0.99and the average absolute error being 6.74 k J/mol,which is equivalent to the experimental error.Based on the prediction model,the importance of each molecular topological index was further ranked,and the relationship between the number of important molecular topological indexes and the performance of the prediction model was also investigated.On this basis,a simplified prediction model for the standard enthalpy of formation was established using the top ten most influential topological indexes.This simplified model remains high prediction accuracy with the R~2 being 0.96and the average absolute error being 20.05 k J/mol.In the aspect of constructing the reaction mechanism,the chemical reaction neural network was used to characterize the differential equation of reaction mechanism.A new training strategy that combines genetic algorithm and stochastic gradient descent method was proposed to solve different characteristic parameters.Based on this strategy,the mechanism of a hypothetical reaction and an alcoholysis reaction in biodiesel preparation was constructed.These two cases show that the step-by-step training strategy can effectively reduce the negative impact of unnormalized data on the training of the chemical reaction neural network.In the aspect of auxiliary engine calibration,based on the engine calibration experimental data,prediction models for parameters in the engine air exchange process were established.For charging efficiency,intake airflow rate,and pumping loss,the determination coefficients R~2 of prediction models were 0.996,0.996,and 0.974,respectively,and the average absolute errors were 0.03,4.98kg/h,and 4.19kPa respectively;for engine torque and power,the determination coefficients R~2 were 0.989and 0.992 respectively,and the average absolute errors were 5.53Nm and2.14kW respectively.Based on these prediction models,the importance of each control parameter to the above output parameters was obtained,and the relationship between the number of experimental data used by the training group and the prediction performance was further explored.The results showed that 1/3 of the current experimental data was sufficient to build a high-precision prediction model,which validates the potential of the active subspace method in simplifying the experimental workload.