Comparison And Research Of Prediction Methods For PM₁₀ Concentration In Shanghai

Inhalable Particulates (PM10) has been the primary pollutant in the air of Shanghai in recent years. The generation, disseminate and transformation of PM10is a nonlinear process with various factors functioning. In order to probe this process, this paper first of all analyzed the temporal and spatial transformation rules of PM10in Shanghai with traditional statistical methods, based on which a forecast model concentrated on the nonlinear approximation capability of modern techniques was established.(1)Analyzing the temporal variations and characteristics of PM10of Shanghai over recent years with statistical methods, it shows that the concentration of PM10in Shanghai decreases year by year over the past decade and has annual variation featuring seasonal fluctuations and diurnal variation featuring double-peak. The distribution characteristics and trends of PM10in different districts and counties of Shanghai are further analyzed demonstrating that the distribution is increasing from outside to inside as a ring. Statistics of the distribution characteristic and trends of in Shanghai and three neighboring cities have done that the concentration of in the three west neighboring cities is higher than Shanghai as a whole. The correlation between API and conventional meteorological data is also calculated that the diffusion and dissemination of PM10is influenced by weather conditions and the influences are seasonal dependent.(2)Using partial least squares analysis combined with cross validation method to extract predictors of API of each season, the results showed that the number of main components were various according to different seasons. Comparing the capability of partial least square regression model and stepwise multiple regression model, it shows that PLS is better than stepwise regression and the PLS is an effective method in sifting predictors.(3) In order to improve the performance of the models, genetic algorithm optimization is used to optimize the initial weights of BP neural network, grid search combined with the cross-validation method is used to optimize the parameters of the SVM model. After examining the performance of both optimized models, we found each has improved to some extent.(4)Based on component matrix sifted from PLS as the forecast factor set, together with the division of four seasons by Meteorology standard, we established forecast models PLS, GA-BP, SVM suitable to each season respectively, and fit the API indicators of last30days of each season in2008. Using four indicators to their forecast performance derived from the comparison and analysis of the statistics, it demonstrates that there are seasonal differences that well performance in summer, while poor in autumn. Since there are differences on each indicator, it remains uncertain to decide the best as well as the worst, it’s suggested to choose the model based on target requirements.