Prediction Of PM_2.5 Concentration Based On Support Vector Regression

Fine particulate matter (PM2.5) has raised broadened public concern these years. Hazy days caused by PM2.5 have largely affected the economy and life of normal people. The prediction of PM2.5 has been a very important task in the control and study of PM2.5. Based on support vector regression (SVR) method, an approach for PM2.5 prediction has been proposed in this thesis. The research is carried out from various aspects concerning PM2.5. The result provides important practical value to the study of PM2.5 and scientific reference to its control.The research in this thesis is listed as follows:(1). Based on the combination of empirical mode decomposition (EMD) and SVR method, the time series of PM2.5 was studied. The EMD method stabilized the PM2.5 raw data and provided decomposition according to varied frequency spectra. Using both false nearest neighbor (FNN) method and SVR, the prediction of PM2.5time series was carried out.(2) A correlation co-efficiency based on SVR method was proposed. The correlation of various parameters with PM2.5 was studied using SVR correlation, by comparison with Pearson correlation and Spearman correlation methods. The result yielded evidence for PM2.5 formation and transformation, and provided instruction for parameter selection in PM2.5 prediction using SVR method. Based on the correlation calculation result, different parameters were selected for SVR model input. Its precision and generalization were tested for PM2.5 prediction in several cities.(3) An enhanced grid search algorithm for super parameter selection of SVR model was proposed based on combination of grid search and simplex algorithm. It was proved that the algorithm was much faster than simple grid search and also comparative with other methods such as genetic algorithm. The method for SVR parameter selection will have its potential in SVR model building and training, An online support vector regression method (OSVR) was proposed for large dataset training in SVR method for PM2.5prediction. The OSVR method effectively reduced the training time for incremental data, especially in the case of large training dataset, and slightly optimized the prediction efficiency of bath SVR process. The OSVR method is highly potential in real application of PM2.5 prediction.