| Along with the larger capacity of power plant and the increased demand of environmental protection,air pollutant emissions from thermal power plant have been under strict supervision.It was difficult to meet the emission standard only by using low-Nitrogen Oxides(NOx)combustion technology,at the same time,secondary flue gas purification must be adopted,including selective catalytic reduction(SCR)flue gas denitration method which has been widely used.As for SCR system,it is important to control NH3 injection.The control effect of NH3 injection was satisfying on stable state.However,when unit operation condition changed,because of complex reaction and interfering factors,SCR system had the features of nonlinearity,great inertia,large delay and time-varying.So,Proportion-Integration-Differentiation(PID)cascade control couldn't guarantee the optimal NH3/NOx mole ratio,what resulted in unsatisfactory control effects.Moreover,as the units of large-scale new energy were integrated into power grid,thermal power units must change the load to improve the efficiency of new energy consumption quickly and extensively.The load fluctuated quickly and widely,so that NOx concentration in flue gas has changed greatly,what was difficult to realize ultra-low NOx emissions.In order to achieve optimal control for NH3 injection,the field data of SCR system was used to establish a data-driven model of outlet NOx concentration,then based on this model,optimal control strategy of NH3 injection was put forward to meet NOx emission standard and ensure denitration efficiency.Meanwhile,the NH3 injection and NH3 escape were reduced.The NH3 injection valve was protected.The stability of control was improved.The main contents:1.Based on the analysis of structure,reaction mechanism and influencing factors of SCR system,the characteristics of the controlled objects and the correlation analysis of the model samples,data-driven modeling of outlet NOx concentration was studied.Based on Kernel Partial Least Squares(KPLS)modeling,considering the characteristics of different kernel function,the multi-scale characteristics of samples and the effect of orthogonal signal correction on model performance,different improved KPLS algorithms were put forward and local predictive model was established.3 UCI datasets and 2 non-linear functions were used to analyze and compare model generalization ability,non-linear fitting ability and anti-noise ability.Considering that the SCR system had large inertia and large time lag was existed in outlet NOx concentration measurement,k-nearest neighbor mutual information(KNNMI)was put forward to estimate delay parameter.Then,based on the analysis of causal relationship among variables,the actual time-delay was confirmed.The phase space of original samples was reconstructed based on the actual time-delay.The verifications of different improved KPLS algorithms showed that outlet NOx concentration could be predicted exactly in advance by using reconstructed samples to model.And,multi-scale wavelet KPLS(mwKPLS)was selected as the modeling based on the analysis of generalization ability and complexity of different improved KPLS algorithms.2.Adaptive approach of outlet NOx concentration prediction model was studied according as SCR system had the characteristics of strong coupling and time-varying.Firstly,according to the need of dynamic modeling for samples,dynamic eliminating outlier and online filtering were put forward.Secondly,due to the interfering factors in reaction process of SCR,bidirectional variable selection was put forward based on Change Rate of KNNMI(KNNMICR)to simplify the model,reduce the calculation during model update process and improve model prediction accuracy.Variable selection experiments on Fridman and Housing data sets showed that compared with other variable selection methods,KNNMICR could select relevant input variables effectively.Field data of SCR system were tested to verify the effectiveness of KNNMICR further based on reconstructed samples.In order to improve the accuracy of model update,delay-time difference(DTD)update strategy and feedback correction strategy were put forward.Compared with the moving window and time difference update strategy,DTD update strategy had the highest accuracy.Feedback correction strategy made that the corrected predicted value was consistent with the change trend of real value.At last,the DTD-mwKPLS prediction model was established by combining mwKPLS algorithm and DTD update strategy.Due to fixed parameters could be adopted to obtain higher prediction accuracy in this model,and the parameter was not required to update frequently.So,time consuming problem caused by massive calculation in model update was avoided.3.In order to compensate response lag of SCR system,make NH3 injection control system more adaptable and protect the NH3 injection valve on variable state,DTD-mwKPLS Model Predictive Control(MPC)was put forward based on Adaptive Particle Swarm Optimization(APSO).Compared with PID,MPC could make NH3 injection valve act in advance,make outlet NOx concentration of the system output could track the set value to avoid repeated oscillation.On steady state and small range of variable state,the accuracy of control and the efficiency of denitration were high.Meanwhile,NH3 injection and NH3 escape were reduced,actuator saturation of NH3 injection valve was avoided.4.In order to improve the control accuracy of PID control and MPC further on large range of variable state,and solve the adverse effect of the inlet NOx concentration measurement lag on system control,NH3 injection composite optimal control strategy was put forward based on hybrid prediction(HP)model of inlet NOx concentration.Firstly,the lag of Continuous Emission Monitoring System(CEMS)measurement was analyzed,then the lag time of inlet NOx concentration measurement was determined based on the change of CEMS blow signal and inlet O2 content signal.Furthermore,the lag time of inlet NOx concentration measurement was confirmed.Secondly,input variables of inlet NOx concentration HP model were selected and preprocessing of CEMS data in blowback process were analyzed,HP model of inlet NOx concentration was established by using exponential prediction model and recursive least square support vector machine,the output of HP model was used as feedforward signal to accurately predict inlet NOx concentration in 60s in advance,then original feedforward control was improved.Finally,combined improved feedforward control with PID control and MPC respectively,NH3 injection composite optimal control strategy was put forward,concluding HPPID control and HP-MPC.The results showed that,compared with PID control,when inlet NOx concentration changed extensively,HP-PID control could make NH3 injection valve act in advance,make the outlet NOx concentration of system output close to the set values.The efficiency of denitration was maintained,NH3 escape was reduced further.But,NH3 injection was easy to overshoot on large range of variable state.Compared with PID control and MPC,HP-MPC could make outlet NOx concentration of system output meet national NOx emission standard,and the data distributed in the range of ±5mg/m3 reached 92.45%,The efficiency of denitration was unchanged,meanwhile,NH3 injection and NH3 escape were further reduced,the action speed of the NH3 injection valve became low,then optimal control of NH3 injection was achieved,and NH3 injection valve was protected. |
PDF Full Text Request