Prediction of ozone hourly concentrations by support vector machine and kernel extreme learning machine using wavelet transformation and partial least squares methods

Abstract

In this paper, we develop a method for predicting ozone (O3) concentration based on kernel extreme learning machine (KELM) and support vector machine regression (SVR) and pretreat it by wavelet transformation (WT) and partial least squares (PLS). To test the method's effectiveness, the observation (2014–2016 summer) of the precursors, meteorology and hourly O3 concentrations in the Nanjing industrial zone were applied. The mean absolute error (MAE), mean absolute percentage error (MAPE), root mean squared error (RMSE), normalized root mean square error (NRMSE) and coefficient of determination (R2) were chosen to evaluate the model. Results demonstrate that the KELM and SVR perform better than stepwise regression (SR) methods and back propagation neural network (BPNN) for predicting O3 concentration. WT decomposes the original time series of O3 concentration into a few sub–series with less variability, and then improve the performance of SVR and KELM by 16.99%~30.91% and 16.00%~25.86%, respectively. The variable importance in projection (VIP) value was used to filter the influence factors of each sub–sequence, which can remove redundant information and reduce the calculation amount of the model. In addition, the WT and PLS methods enhance the predictive abilities of KELM and SVR for higher O3 concentrations by 21% and 35% respectively. The KELM-WT-PLS model shows the best fit of the O3 hourly concentration, and the corresponding MAE, MAPE, RMSE, NRMSE and R2 are 7.71 ppb, 0.37, 9.75 ppb, 11.83% and 0.78, while KELM predict the O3 hourly concentration more accurately.