A hybrid-wavelet model applied for forecasting PM2.5 concentrations in Taiyuan city, China

Abstract

The main objective of this study is to apply hybrid-wavelet models developed by combining artificial intelligence (AI) techniques and wavelet transforms to the prediction of PM2.5 concentrations. As a representative of energy and industrial cities in Northern China, Taiyuan is taken under consideration in this study. Hourly data of PM2.5 concentrations of two years are decomposed with wavelet transforms including db4, db6, db9, coif4, haar, sym3 and dmey. Then, the high-dimensional input vectors of AI models are constructed with factors selected by neighborhood mutual information (NMI) which are highly correlated with PM2.5 concentrations. In the hybrid models, the AI models are back propagation neural network (BPNN) models representing the artificial neural network (ANN) model and support vector regression (SVR) models representing the support vector machine (SVM) regression algorithm. To evaluate the model accuracy, the results obtained by these hybrid-wavelet models are compared with the results of single ANN and SVM. A comparison of the performance indicators of the models indicates that the results of the hybrid models are more accurate and stable. The MAE (mean absolute error) of the best ANN-wavelet model with regard to the 2016 ANN model decreased by 39.79% and in the optimized case for 2017, it decreased by 41.95%. The most accurate SVM-wavelet model had an MAE that decreased from 10.8347 to 5.7195 to 9.4041 and 5.1401 respectively for the data of 2016 and 2017.