Prediction And Space Distribution Analysis Of PM2.5 Daily Average Pollution Levels Based On ESN Model In Beijing

The monitoring and prediction of PM2.5 pollution level and the spatial distribution characteristics related to the human health, animal and plant growth, assessment of atmospheric environment and climate conditions analysis and so on. In the field of atmospheric environment, PM2.5 pollution has gradually become one of the great issue to many people.This paper was based on the daily average PM2.5 pollution concentration data and the wind speed and wind direction data in January, April, June and October of 2013, taking Beijing City as the study area, using echo state network model to predict each month daily average PM2.5 pollution concentration respectively. Combined with geostatistics method, using semi variogram to analysis the spatial variation characteristics of PM2.5 daily average concentrations. On this basis, the measured values and predicted values were interpolated with the Kriging interpolation method. At last, the work compared the spatial distribution of similarities and differences respectively. In this study the main work and conclusions are summarized as follows:i. The ESN neural network model and traditional BP neural network were used to forecast Beijing daily average PM2.5 concentration time series respectively. The compared results demonstrate that ESN model has higher prediction accuracy. Among them, the correlation coefficient in January was 0.6977; the correlation coefficient in April was 0.8924; the correlation coefficient in June was 0.663; the correlation coefficient in October is 0.7946.ii. The ranges of PM2.5daily average concentration data in the four months were all large. This indicated that the variation range of the spatial autocorrelation of PM2.5 daily average concentration were large. The nuggets had large differences in different time of the study area. The values of these nuggets roughly between 11.1 and 26.1 which shows that the random factors, system error have a great effect on the spatial autocorrelation of PM2.5 daily average concentration. The nugget-to-sill ratio value were between 0.08 and 0.31, which indicated they have strong autocorrelation.iii. Based on the results of kriging interpolation, classification statistics was applied according to the classification standard of pollution level. Spatial distribution between measured and predicted values in January are divided into two levels: high levels of pollution and serious pollution. The graded distributions have the similar trend. April hierarchical distribution of predicted and actual values emerged:Good. The overall classification distribution of the predicted and measured value in June was belongs to light pollution. In grading distribution of the predicted value, moderately polluted section contains small area of Chaoyang District and larger area of Shunyi District; the air quality in region centered on the junction of Yanqing District, Changping District, Huairou District was good. In October, the grading distribution of the predicted and measured values are divided into two levels: good and light pollution. The overall pollution tendency is similar with good air quality distribution in northern, light pollution in southern.iv. Combined with GIS spatial analysis, the result shows that the spatial distributions of the measured and predicted values have high similarity. Therefore, the application of ESN network model PM2.5 pollution level time series of single step prediction in Beijing is feasible.