Modeling Method Of Wet Desulfurization Process Based On Neural Network Technology

The limestone-gypsum wet flue gas desulfurization process for coal-fired power plants is the main desulfurization process at present because of its mature technology,significant desulfurization effect and high reusability compared with other desulfurization processes.Wet flue gas desulfurization is a multi-variable,strongly coupled,non-linear and complex process.It is crucial to coordinate the dynamic balance among the factors to improve the efficiency of the flue gas desulfurization system.Modeling,optimizing and predicting the outlet flue gas SO₂ content of the flue gas desulfurization system and thus achieving higher flue gas desulfurization efficiency is of practical importance in terms of the operating cost of the power plant.This thesis makes the following main works.（1）Analysis of limestone-gypsum wet desulfurization process.Research and analysis of the limestone-gypsum wet flue gas desulfurization process focuses on the core equipment,the flue gas desulfurization absorption tower,analyzing its internal reaction to absorb SO₂ in the flue gas principle,deriving the various factors that mainly affect the SO₂ content at the outlet,and determining the information on the data variables to be collected.The required neural network theoretical basis is also listed.（2）Data pre-processing and the neural network model is built and trained.To facilitate model training and improve the accuracy of training,the data collected in the field are preprocessed.Box plots are used to process the abnormal data accordingly,and to ensure the temporal order on the data,filling by mean-fill method.After the amount of data are all in a complete and reasonable state,data downscaling is performed and Pearson correlation coefficient is used to find out the variables with strong correlation with the exported SO₂ content as the input variables of the model.Finally,wavelet threshold noise reduction is applied to all the data,and the obtained data are smoother and can represent the values of various indexes at stable working conditions.The study constructs a modified long and short-term memory neural network,including two long and short-term memory layers,two linear rectifier function layers,a fully connected layer and input and output layers.The input variables of the model are raw flue gas SO₂ content,p H,slurry density,slurry flow,flue gas temperature and SO₂content of the exit flue gas at the previous moment,and the output variable is the SO₂content of the exit flue gas.The model in this thesis is analyzed by its own comparison and comparison with BP neural network,recurrent neural network and basic long and short-term memory neural network,and the accuracy of the model built in this thesis reaches 97.7%within the error range of[-1,1],with the loss values of 0.0033 and0.00581 during training and testing,respectively.（3）The desulfurization system process is built.Make a detailed introduction to the actual projects involved,make targeted adjustments to the problems that exist in the actual working conditions on site,and the site finally adopts a common control mode of one or two levels.The secondary control system mainly realized the construction of the operation screen and the control script writing by using C#.Huaneng Yingkou Power Plant’s#4 unit desulfurization system was used and achieved significant results.