论文部分内容阅读
We developed and tested an improved neural network to predict the average concentration of PM10 (particulate matter with diameter smaller than 10 μm) several hours in advance in summer in Beijing. A genetic algorithm optimization procedure for optimizing initial weights and thresholds of the neural network was also evaluated. This research was based upon the PM10 data from seven monitoring sites in Beijing urban region and meteorological observation data, which were recorded every 3 h during summer of 2002. Two neural network models were developed. Model I was built for predicting PM10 concentrations 3 h in advance while Model II for one day in advance. The predictions of both models were found to be consistent with observations. Percent errors in forecasting the numerical value were about 20%. This brings us to the conclusion that short-term fluctuations of PM10 concentrations in Beijing urban region in summer are to a large extent driven by meteorological conditions. Moreover, the predicted results of Model II were compared with the ones provided by the Models-3 Community Multiscale Air Quality (CMAQ) modeling system. The mean relative errors of both models were 0.21 and 0.26, respectively. The performance of the neural network model was similar to numerical models, when applied to short-time prediction of PM10 concentration.