Research On Data-driven Modeling And Compound Optimal Control For SCR Flue Gas Denitration System In Coal-fired Power Plant

At present,thermal power generation,especially coal-fired units,will still occupy the main position in China’s power structure.And the environmental pollution problems caused by coal-fired units cannot be ignored.In order to achieve ultra-low pollutant emissions,the existing coal-fired units in China are equipped with tail flue gas denitrification systems,of which the most widely used is the Selective Catalytic Reduction(SCR)denitration system.Nowadays,the NH3 injection control system in the SCR denitration system mainly adopts the Proportional Integral Differential(PID)cascade control mode,which improves denitration efficiency by adjusting the NH3 injection flow rate.However,due to the nonlinear and large lag characteristics of the SCR denitrification reaction process,it is difficult for NH3 injection to quickly adapt to the change in the inlet Nitrogen Oxide(NOx)concentration.When the amount of NH3 injection is too low,it is easy to lead to the reduction of denitration efficiency and NOx emissions exceed the standard.And when the amount of NH3 injection is too high,generates by-products that reduce the catalyst activity and block the air preheater.In addition,excess NH3 will also corrode the air preheater,resulting in waste of ammonia cost and secondary environmental pollution.In the future,as a large number of new energy power sources are integrated into the power grid,coal-fired units will operate under low or variable load conditions the most of time,thus increasing the difficulty of achieving ultra-low NOx emission of coal-fired units.Therefore,it is necessary to optimize the existing NH3 injection control system.In order to better realize the optimization of the NH3 injection control system,it is necessary to establish a data-driven model of the SCR denitrification system as accurately as possible.The field operation data usually show nonlinearity,hysteresis,diversity,and other characteristics that are not conducive to modeling.Therefore,it is first necessary to improve the traditional data-driven model to improve the generalization ability of the model,and optimize the existing NH3 injection control system to achieve more precise control of the amount of NH3 injection,maintain a high denitration efficiency,and minimize the NH3 escape.It is of great significance to ensure the safe operation of the unit and reduce the operation cost of denitrification.This paper analyzes the characteristics of the combustion system and SCR denitration system of coal-fired units,as well as the problems existing in the existing NH3 injection control system,and studies three aspects:data preprocessing data-driven model improvement,and NH3 injection control optimization.(1)A outlier detection method based on Particle Swarm Optimization(PSO)algorithm to optimize the Probabilistic Neural Network(PNN)(PSO-PNN)is proposed to identify outliers in the historical data of the SCR denitration system.At the same time,the Nadaraya-Watson regression multiple imputation method based on sliding window optimization is proposed to fill in the abnormal missing values of the historical data of the SCR denitration system.The simulation results show that the PSO-PNN method has good outlier detection performance,and the Nadaraya-Watson regression multiple imputation method based on sliding window optimization can achieve a better thermal data-adaptive filling effect,which can meet the needs of data preprocessing for subsequent modeling.(2)An algorithm based on Ensemble Empirical Mode Decomposition(EEMD)and Improved Whale Optimization Algorithm(IWOA)to optimize Deep Extreme Learning Machine(DELM)is proposed to establish a data-driven prediction model for inlet NOx concentration at the SCR denitration system.The simulation results show that the model combined with EEMD technology has the advantages of being able to handle nonlinear and non-stationary signals,making the model suitable for a wide range of variable working conditions and improving the accuracy of the prediction model;and IWOA is used to optimize the parameters of the DELM model,and then the combination of EEMD and IWOA-DELM forms complementary advantages and further improves the prediction accuracy of the model.(3)It is proposed to introduce the attention mechanism into the Convolutional Neural Network Bidirectional Gated Recurrent Unit(CNN-BiGRU)to establish a data-driven prediction model of outlet NOx concentration at the SCR denitration system.The simulation results show that the use of BiGRU can learn the dynamic changes of the time series constructed by the feature vector of CNN,realize dynamic data-driven modeling,and improve the generalization ability of the model.and the introduction of attention mechanism can give BiGRU the different probability weights of the hidden layer states by mapping weight and learning parameter matrix to reduce the loss of historical information,strengthen the influence of important information on the outlet NOx concentration,and further improve the prediction accuracy of the model.(4)The compound optimal control of the SCR denitration system is proposed that the output of inlet NOx concentration prediction model at the SCR denitration system is used as a feedforward signal,and the output of outlet NOx concentration prediction model at the SCR denitration system is passed through the differential controller,the change of the outlet NOx concentration is feedback to the NH3 injection control system in advance.The simulation results show that the outlet NOx concentration is always lower than the emission standard by optimal control,which effectively reduces the NH3 escape,keeps a high denitration efficiency,and avoids the saturation of the valve actuator.At the same time,the control strategy is applied in the field,a better control effect was also achieved.