Search

PM 10 is the main air pollutant in Taiyuan, as the city is a heavy industrial center with coal as its main energy source. Therefore, research on the prediction of this pollutant's variation and concentration is of great theoretical significance for air pollution prevention and emergency solutions. The source of PM 10 is very complex, as it is affected by industrial emissions, vehicle exhaust, fugitive dust, and many other factors. The emission sources of PM 10 are difficult to determine accurately. The goal of our research was to give accurate forecasting results efficiently when only time-series PM 10 concentrations, and no other exogenous information, is available. A support vector machine (SVM) enjoys good generalization performance in the PM 10 concentration forecasting area. Traditionally, an SVM chooses historical data as the input features in the process of dealing with the time-series data of air pollutant concentrations. However, data with simple structure and incomplete information have become the fetter of generalization ability improvement. In this study, the data for simulation experiments was the PM 10 concentration dataset collected from four monitoring stations in Taiyuan. The PM 10 concentration time-series one-dimension data was decomposed into high dimension, constructed by low frequency and high frequency series using a wavelet transform. The wavelet-SVM forecasting model can be established by introducing the high-dimension data as the input features. The experiment results indicate that, contrasted with the traditional SVM, the wavelet-SVM model boasts higher accuracy for PM 10 concentration prediction. In particular, it captures the concentration mutational points more accurately and provides information support that is more effective for atmospheric pollution warning. In addition, with the wavelet-SVM model, prediction accuracy for the concentration variations was significantly improved and laws that were more inherent in the PM 10 concentration time series were revealed.