Research And Implementation Of Soft Sensor And Prediction Concerning Emission Concentration In Coal-Fired Power Plants Based On Deep Neural Network

Electricity has become an indispensable energy source.Although there are various power generation methods,coal-fired power generation accounts for the highest proportion.In the process of coal-fired power generation,coal combustion will produce a large amount of polluting gas,and NOx are the most harmful to the environment.The new trend of production is energy-saving and environmentally friendly.Environmental protection department of my country has issued relevant documents to put forward clear and strict requirements for the emission of NOx from coal-fired boilers in different regions.In order to reduce NOx emission in flue gas,SCR flue gas denitrification technology is adopted by most power plant enterprises because of its high denitrification efficiency and low maintenance cost.The principle of flue gas denitration technology is:Ammonia gas reduces NOx to non-toxic and harmless nitrogen and water vapor under the action of catalyst in the denitration reactor.When the ammonia injection is insufficient in the denitration process,NOx cannot be completely reduced.The by-products generated by excessive ammonia injection during the denitration process will not only reduce the denitration efficiency but also corrode downstream equipment.In order to reasonably control the amount of ammonia injection,it is necessary to accurately predict the NOx concentration in the inlet flue gas of the SCR destocking reactor.This paper takes a domestic coal-fired power boiler as the research object,and the contributions are as follows:(1)In this paper,the mechanism of NOx generation in the flue gas of coal-fired power plants is analyzed.We analyze NOx flue gas control technologies commonly used in power plants.In this paper,the reaction process of SCR denitration reactor,the formation of byproducts and the influencing factors of NOx generation are studied,which lays the foundation for the establishment of NOx soft sensor model and prediction model.Based on the time series data of the DCS system of Luoyang Coal-fired Power Plant,the data characteristics of the thermal system are analyzed.Therefore,it is very challenging and difficult to predict flue gas NOx based on the time series data of the thermal system of the power plant.The soft sensor modeling and prediction process of NOx in the flue gas of the denitrification reactor is analyzed for the thermal system of the power plant.(2)A soft sensor model of hybrid structure neural network based on deep learning is proposed.One branch is used to learn the sequence data features,and the other branch is used to learn the increase and decrease trend of the NOx parameter.In order to realize the real-time guidance of the soft sensor model for ammonia injection in the denitrification reactor,an online time series data acquisition software based on the InfluxDB database was implemented.In the model learning process,the weight structure of the hybrid network is dynamically adjusted according to the epoch in the training process,so that the NOx soft sensor model has a more accurate trend prediction accuracy on the basis of small errors.The hybrid network structure improves the robustness and generalization performance of the soft sensor model,and has important guiding significance for accurate ammonia injection in denitrification reactors.(3)A neural network structure model is proposed,which uses wavelet transform to denoise time series data,uses stacked autoencoder to extract data features,and then uses GRU to predict NOx concentration.The optimal layer structure of the stacked autoencoder is determined experimentally,the neural network structure has better performance on the task-driven reorganization dataset based on average mutual information.Experiments show that the designed network structure has strong ability of time series data feature extraction and high NOx prediction accuracy.In order to realize the application of the prediction model and the soft sensor model in the actual working environment,this paper transforms the model and compiles the dynamic link library,which can facilitate the training of the model and apply it to the power plant system to provide guidance for ammonia injection.