Prediction of PM2.5 concentration based on improved secondary decomposition and CSA-KELM

Abstract

Exposure to high concentration PM2.5 will increase the risk of human illness and death, so it is of great significance to establish a high-accuracy PM2.5 prediction model. Because PM2.5 concentration sequence is nonlinear, non-stationary and complex, how to accurately predict PM2.5 concentration becomes a difficult problem. To improve the prediction accuracy, a PM2.5 prediction model based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), differential symbolic entropy (DSE), variational mode decomposition improved by butterfly optimization algorithm (BVMD) and kernel extreme learning machine optimized by crow search algorithm (CSA-KELM), named CEEMDAN-DSE-BVMD-CSA-KELM, is proposed. The idea of CEEMDAN-DSE-BVMD-CSA-KELM is roughly as follows. Firstly, CEEMDAN is selected as the main decomposition technology, and PM2.5 concentration sequence is decomposed into several IMFs and residual term. Calculate the DSE of each component to judge the complexity and find out the high complexity components. Secondly, BVMD decomposes the sum of the high complexity components again, and obtains several new components. Finally, CSA-KELM is used to predict the residual components of primary decomposition and the components of secondary decomposition, and the prediction is completed by summing up the component prediction results. To prove the superiority of the proposed model, it is compared with single prediction model, single multi-factor prediction model and other combined prediction models. Taking the PM2.5 concentrations in Beijing, Shenyang and Shanghai from January 1, 2016 to March 31, 2021 as examples, the simulation experiment is carried out by using MATLAB platform. The fitting coefficient R2 of each city is as high as 0.99, and the RMSE, MAE and MAPE values of Shanghai are 0.7911, 0.5978 and 0.0167, which are lower than those of other comparison models. This proves the proposed model has high prediction accuracy and stability.