Reconstruction Of Pulverized Coal Combustion Process Based On CFD And Machine Learning

The Chinese strategy of carbon neutralization and carbon peak requires coal-fired boilers in power plants to flexibly adjust the working conditions in the face of complex power grid load changes,such as adjusting the amount of coal supplied by coal-fired boilers or the number of burners working.However,the frequent adjustment of boiler operation has brought challenges to the stable operation of boiler combustion.Reproducing the combustion status of the boiler under various working conditions can provide reference for the on-site operators,so as to better guide the operation of the coal-fired boiler and make it meet the requirements of safety,energy saving and environmental protection.In this paper,the physical and chemical composition field of boiler furnace is quickly and accurately reproduced by using running data and temperature measuring points near the burner.Firstly,a computational fluid dynamics model was established based on a 400t/h tangential boiler,and the initial conditions were set according to the actual operation of the boiler.The numerical simulation of its combustion process was carried out by Fluent software,and the accuracy and reliability of the numerical simulation of the boiler were verified.Then,by adjusting the coal feed of the bottom burner and the corresponding secondary wind speed,60 numerical simulation results are generated.The temperature,velocity and chemical composition data of the bottom burner plane and vertical plane were extracted as data sources for the subsequent deep learning reconstruction.Secondly,it uses Transpose Convolution Neural Network(TCNN)based model to learn the distribution of temperature field,velocity field and chemical composition field of key sections in the combustion region under different working conditions.In particular,the working condition data and temperature value detected by burner fire are taken as neural network input.The temperature field,velocity field and chemical composition field of the plane where the bottom burner is located and the vertical section of the furnace are taken as output.The results of neural network reconstruction show that the average prediction error of temperature field for the plane where the bottom burner is located is as low as 2.56%,the prediction error of velocity field is as low as 4.71%,and the correlation index of carbon monoxide and oxygen concentration prediction results and numerical simulation results are 0.932 and 0.978,respectively.In terms of time cost,The TCNN network training time is two orders of magnitude less than the numerical simulation calculation time,and the network prediction time is almost negligible,achieving the goal of fast and accurate reconstruction of boiler furnace parameter field.The reconstructed temperature and velocity fields of vertical section are predicted well,and the universality of the reconstructed network is verified.In addition,the experimental comparison with the deep confidence network model shows that the TCNN network has better reconstruction accuracy.In order to verify the effectiveness of introducing the burner fire detection temperature value,the combination of different number of fire detection temperature points is discussed.The results show that with the increase of the number of burner fire detection temperature points,The average error of TCNN’s prediction of the next wind plane temperature field gradually decreased from 11.14%to 2.56%,and the prediction effect of velocity field and product concentration field was also significantly improved.The experiment showed that the introduction of fire temperature detection value effectively improved the accuracy of the reconstruction algorithm.The network hyperparameters in TCNN were compared,and different network layers,batch sizes and learning rates were compared to determine the network hyperparameters of optimal reconstruction accuracy and minimum training time.