AQI Prediction Based On LASSO-ELM Nonlinear Combination Model

With the rapid development of economy,air pollution has become a key concern of the whole society.In recent years,haze has appeared in most parts of China,and it was one of China’s national security events from 2013 to 2017.Air pollution prevention and control is an important way to eliminate haze,and air quality prediction can provide predictive information for air pollution prevention and people’s travel.Therefore,it is of great practical significance to establish a scientific and effective air quality prediction model.From the existed researches,this thesis proposes a CEEMD-LASSO-ELM combined model to provide technical support for reliable AQI prediction.Combined prediction is a popular statistical method for air pollution forecasting,which generally includes two steps of individual models selection and weight determination.Individual models selection is the first step for combined prediction,which also directly affects the results of combined prediction.Reviewing the literatures,it can be found that many combined prediction researches did not select individual models,but blindly established combined prediction,which led to inferior combinations.For the traditional combined prediction theory,the weights of the linear combination are generally optimized based on the restriction of non-negative weights summed to 1.Some scholars also point out that the weights limit can be extended to general linear regression.In fact,if improve the accuracy be the goal for combined prediction,we should not stick to the linear combination,and explore the nonlinear combined prediction.To solve the above problems,this thesis introduces LASSO for individual models selection,and ELM to establish a nonlinear combined prediction model,named CEEMD-LASSO-ELM.LASSO can not only select individual models,but also avoid the possible collinearity between the selected individual models.The specific steps of CEEMD-LASSO-ELM are as follows: First,CEEMD is used to decompose the original AQI to some IMFs,which are combined into four modes according to the frequency characteristics.Second,PSOGSASVR,GRNN,CNN and LSTM are used to simulate and predict each mode respectively.The modes’ predictions are added up to obtain 256 individual models.The original AQI sequence regresses on the 256 individual models by LASSO,and the individuals are selected for combined model.Finally,the ELM is employed to build the nonlinear combined prediction for AQI.In this paper,daily AQI series of Guangzhou,Kunming,Lanzhou and Hulunbuir are selected to verify the feasibility and effectiveness of the proposed CEEMDLASSO-ELM model.Combined with CEEMD,seven comparative models are set to be Random-ELM,Rank-ELM,Random-LR,Rank-LR,LASSO-GRNN,LASSOPSOGSASVR and LASSO-LR.The data analysis shows that the proposed CEEMDLASSO-ELM model has better prediction accuracy and stronger generalization ability.Taking Lanzhou for example,the error MAPE of CEEMD-LASSO-ELM model is2.154% lower than that of the seven comparative models on average.