Research On Ozone Concentration Prediction And Method Comparison Based On Kernel Extreme Learning Machine And Wavelet Transform

The generation and dissipation of near-ground ozone?O₃?is a non-linear and strongly coupled process.In recent years,with the development of computer technology,artificial intelligence technologies such as machine learning have shown certain advantages in solving nonlinear prediction problems.Among them,the extreme learning machine with kernel?KELM?,as a newer machine learning method,has not been fully applied in the field of atmospheric environment.This study was based on the O₃,precursors,and meteorological data observed in the northern suburbs industrial zone of Nanjing during the summer of 2014-2016.this study mainly carried out the following two aspects:First,based on the regression function of KELM,the prediction method of O₃ hourly concentration in the northern suburb of Nanjing in summer was established and compared with stepwise regression?SR?,support vector machine regression?SVR?,and back propagation neural network?BPNN?.Moreover,the application effect of KELM classification function on the problem of O₃ excessive concentration was explored.The results indicated that the KELM showed the best prediction performance among the three machine learning methods in the regression prediction problem,with the advantages of high prediction accuracy and fast learning speed.The MAE,RMSE,NRMSE,R²,and IA between the KELM predicted value and the observed value WERE 10.43 ppb,13.48 ppb,16.18%,0.64,and 0.86.The performance of SVR and BPNN on prediction problems was not much different,while SVR has a slight advantage in RMSE,NRMSE,IA and learning speed.In the problem of classification and recognition,the accuracy of KELM-C1 in early warning of high hourly concentration of O₃ was90.63%,slightly higher than that of SVC1.All"normal"conditions were correctly judged.The accuracy of KELM-C2 in early warning of maximum 8-hour average concentration of O₃ was82%,which was not much different from SVC2.Second,the wavelet transform?WT?and VIP value based on partial least squares method?PLS-VIP?were used to optimize the KELM prediction method,and compared it with SVR.The results showed that WT method decomposed the original O₃ time series with high variability into several subsequences with low variability,which effectively improved the accuracy of O₃ concentration prediction by 2.33%?21.82%.PLS-VIP method was used for variable screening at various levels,which reduced the number of input variables from 14 to 2-6 while ensuring prediction accuracy,and shortened running time of the model.The two pretreatment methods,WT and PLS-VIP,effectively enhanced the ability of the original model to predict different concentration range of O₃,especially high-concentration which was the focus of air pollution prediction.After overall comparison,the prediction effect of the models based on KELM was 5.91%better than the models based on SVR,which had a good application value.