| With the continuous improvement of China’s development level and environmental protection concept,the prevention and control of air pollutants has gradually become the key point of thermal power upgrading.NOx’s strong pollution and large emission is the focus of prevention and control of coal-fired power stations.This paper will establish a prediction model of NOx at SCR inlet.Simulating the change law of NOx mass concentration from the perspective of data intelligent learning is conducive to the accurate control of NOx concentration by the power station system.The model is based on a 660 MW tangential coal-fired unit.First of all,a total of 125 characteristic variables related to NOx mass concentration are selected from five categories from the three aspects of air,powder and furnace.According to the uneven characteristics of the power station data,the segmented cubic Hermite interpolation is used as the time series normalization method to unify the data time series on the basis of ensuring the data continuity.The abnormal data is filtered through the Laida criterion and the box diagram to improve the data quality.Distinguish the unit status of the data based on the sliding pane method.After preprocessing steps such as time series normalization,data cleaning and steady-state extraction,the formed data has a standardized format,which significantly improves the data quality.The formed standardized data set consists of 125 variables as a group,with a total of 259200 groups of data.The calculation results based on grey correlation analysis and CRAT regression tree show that the sensitivity analysis method can not accurately judge the correlation degree of each parameter and NOx concentration in the traditional coordinate system.Moreover,the internal differences of the correlation coefficient sets obtained by each algorithm are small,and it is unable to effectively extract low correlation variables.Based on principal component analysis(PCA),the current coordinate system is linearly transformed.When the feature set contains 50 principal components,the information content has reached 99.24%of the original number set.Therefore.although PCA lost some information,it reduced the NOx mass concentration prediction set from 122 variables to 50 principal components.greatly reducing the calculation amount of the model.According to the calculation results,a feature set reduction strategy including physical mechanism reduction,filtering reduction and PCA dimension reduction is finally formed.The use of extreme learning machine(ELM)as a training algorithm reduces the search space while still having strong generalization ability.Combining particle swarm optimization(PSO)to further improve ELM accuracy.Finally,a prediction model of NOx mass concentration at SCR inlet based on PCA-PSO-ELM was established.Pre-debug the key parameters of the PSO algorithm according to the Rastigrin function.While maintaining the local convergence rate of PSO,the convergence algebra of PSO is reduced to 29.96.Compared with other single hidden layer neural networks,the goodness of fit(R2)of PCA-PSO-ELM reached 0.8957.While ensuring the training accuracy,the training time is at least 72.85%shorter than other methods.PCA-PSO-ELM was used to predict the data of different unit states,and the prediction accuracy of the model for variable load reached 98.43%and 99.14%of the stable load.The model can adapt to the impact of variable working conditions on data,and has good robustness and universality.Establish zones for prediction based on different oxygen levels and loads,the prediction accuracy of oxygen-enriched combustion zone and low NOx combustion zone is 1.76%and 4.66%higher than that of full-load prediction respectively.The zoning prediction method based on PCA-PSO-ELM can quickly,efficiently and accurately predict NOx mass concentration and adapt to different unit conditions,which is conducive to guiding the accurate ammonia injection of SCR system under complex conditions. |
