CEEMD-subset-OASVR-GRNN for ozone forecasting: Xiamen and Harbin as cases

Tropospheric ozone is a standard air pollutant, and can adversely affect human respiratory system. Many metropolitan areas around the world struggle to meet ozone standards. Therefore, timely and effective ozone prediction can help regulatory agencies to prevent the harm to human body and environment induced by excessive ozone concentration in advance. For selecting the optimal individual model set in combined forecasting, this research proposes CEEMD-Subset-OASVR-GRNN model, based on complete ensemble empirical mode decomposition (CEEMD) method, support vector regression (SVR), generalized regression neural network (GRNN) and optimization algorithms (OA), to predict the daily average concentration of ozone. Specifically, for the ozone time series, CEEMD is used to decompose the original data into three intrinsic mode functions (IMFs), PSO-SVR, PSOGSA-SVR, GWO-SVR and GRNN are employed to model and predict the IMFs, and the prediction results are randomly combined to establish 100 individual models (appendix Table A1). The selection methods of the individual models include MSE ranking, factor score and unsupervised learning systematic clustering, and the influence of number of individual models on combined forecasting is studied. The ozone series of two very distinct Chinese metropolitan areas, Xiamen and Harbin, are selected as the experimental data. The prediction results show that the systematic clustering method is helpful for effectively improving the prediction accuracy of the combined model.