Research On No_x Concentration Prediction Method Of Thermal Power Plant Based On Improved Stacked Autoencoder

Energy is an important resource in economic and social development,and fossil fuel is still the main form of energy in China.In China,about 70%of electricity is generated by thermal power plants.However,coal combustion will generate NO_x.If NO_x is not eliminated,it will cause environmental pollution and harm human health.NO_x elimination is inseparable from the prediction of its emission concentration.Once a suitable NO_x emission prediction model is established,NO_x can be reasonably eliminated to minimize the impact of emission pollution and minimize the economic cost while protecting the environment.Therefore,it is very important to establish an accurate NO_x emission prediction model for environmental control and safe and economic operation of thermal power plants.However,NO_x generation and emission process which is influenced by many factors is complex in the actual thermal power plant,and direct modeling is not easy.With the explosive growth of data in boiler combustion system of thermal power plant,researchers pay more and more attention to the data-driven method,especially the outstanding advantages of deep learning in data mining and non-linear modeling,which makes it the most concerned soft sensor method in the era of big data.The existing NO_x prediction results are obtained under the assumption that the data obey Gaussian distribution,but in fact,this assumption is difficult to meet.Therefore,this paper describes the non-Gaussian characteristics based on the statistical information criterion,and uses the improved stacked Autoencoder to establish the NO_xconcentration prediction model.The specific research contents are as follows.Aiming at the nonlinear and non-Gaussian characteristics of NO_x emission concentration in thermal power plants,this paper proposes a generalized correntropy based gated stacked target-related Autoencoder(GC-GSTAE),which uses deep learning method to establish nonlinear model,and introduces loss function based on high-order statistics to describe non-Gaussian characteristics of data.By comparing with loss function based on mean square error,it is proved that the performance index describing non-Gaussian characteristics of data is better and the prediction accuracy is higher when predicting NO_x emission concentration of thermal power plant.Aiming at the problem that the number of hidden layers is too many in deep learning,which leads to the decline of accuracy,this paper proposes a generalized correntropy based gated stacked input-output-related Autoencoder(GC-GSIOAE).When extracting features from hidden layer,it not only learns features from the previous hidden layer,but also learns features from the original input,and fuses different forms of information to ensure the integrity of information.In the simulation comparison,a numerical simulation example is used to illustrate,and then GC-GSIOAE is applied to the NO_x concentration prediction of thermal power plant,which proves that this method can improve the prediction accuracy.Aiming at the dynamic characteristics of NO_x emission in thermal power plants,a generalized correntropy based stacked gated recurrent unit Autoencoder(GC-SGRUAE)algorithm is proposed.NO_x emission of thermal power plant does not only depend on the operating conditions at the current moment,but also the operating parameters at the previous moment will affect the emission at the current moment or even at several subsequent moments.GC-SGRUAE algorithm can not only deal with the prediction of dynamic data and analyze the time correlation between information,but also ensure that gradient explosion or disappearance does not occur easily while deepening the network depth,which is more conducive to mining the inherent abstract characteristics of data.Through the simulation analysis of thermal power plant data,the effectiveness of GC-SGRUAE algorithm in NO_x concentration prediction is verified.