Neural network model of industrial plant’s harmful emissions

Abstract

In many cases, direct measurements of physical quantities are not available. To solve the problem it is possible to create hybrid models of systems or processes based on artificial intelligence (AI). Such models consist of two parts: the first model is based on physical laws and the second one is built on a neural network (NN). The first model generates training samples for the second model. In this study, a hybrid model of the harmful emission (NOx and CO) from the combustion chamber of an aircraft engine was developed. The first stochastic nonlinear mathematical model is grounded on the Zeldovich equation. It applies the superposition principle of oxides’ production in diffusion and homogeneous flames. The basis of the Zeldovich-technique is the probability density distribution functions of the concentration of fuel-air composition in the homogeneous burners respect to a spatial heterogeneity of the mixture composition and a harmonic component of the acoustic waves generated by the heat release. The second model is based on the neural network. The example of a neural network and the results of its training on a real industrial gas turbine combustion chamber are presented. It is shown that the two-layer neural network having 10-35 neurons in the hidden layer provides an acceptable error (not exceeding 5%) of the NOx and CO emission estimation and can be used as a model of harmful emissions to the atmosphere.