Neural Network Prediction Study On SO₂ Concentration Of Suzhou City

IMET(Italian Ministry for the Environment and Territory ) and SEPA(State Environmental Protection Administration of China) set up a Air Quality Monitoring System (AQMS) in Suzhou city, in 2001. To make full use of AQMS data to assess and predict the air quality in Suzhou city, IMET, EMC and INET started up a co-operation, a study on prediction of air quality in suzhou city, utilizing neural network(NN) method.During the preprocessing work, air stability and mixing layer height data were derived from original data of 2004. Wind direction data is transformed to weighted input of the network. Then correlations of the atmosphere factors were analyzed, in time domain and of the mutual influence. According to the physic model of SO₂ diffusion in boundary layer, 13 factors, which have high correlation with SO₂ , were chosen to be the input array of the network.After a research on the fundamental of the artificial network, we set up a 3-layers MLP network structure. We did mass calculation work to figure out the performance curve, which effected by hidden-layer neurons number of NN, and then optimized the neurons number. To get better performance, we also compare 9 different BP-based training algorithms, and the CGP algorithm is proved best. Test by data set of 2004, the well-trained NN works well, the mean error of year is less than 13%.To find out the atmosphere factors'weight of effect on SO₂ , we analyzed the sensitivity of different factors, and it helped to know more about the cause of SO₂ pollution of Suzhou city.For convenience, we re-built the network by C++ language, using the C language interface provided by MATLAB. And we also built a user interface in windows platform .