Research On Air Pollution Prediction Technology Based On Neural Networks

With China's rapid economic development,the process of modernization caused great pressure to the atmosphere environment.Currently,the air pollution problems of sulfur dioxide,nitrogen oxides,particulate matter etc.are getting more serious.The serious air pollution even becomes a great potential threat to the "Green Youth Olympic Games" in Nanjing in 2014.In order to better grasp the variation of atmospheric pollutants,take earlier preventive measures to air pollution events,and cast more targeted control and prevention of pollution,air pollution prediction research is of great significance.Currently air pollution forecasting is based on statistical forecasting and numerical prediction.China is in use of the statistical prediction method.Based on pollution emissions,meteorology and air quality,this method relies on massive data,is not sensitive to air changes and predicts in low accuracy.The numerical prediction is based on the atmospheric model.This method has a huge structure in need for complex dynamics,physics,and chemistry theories.It requires massive weather conditions,pollution emissions and urban terrain underlying surface data,thus make it difficult to predict and time-consuming to research and develop.The main work of this paper includes the following points:1)We propose a pollutant prediction model based on Back Propagation neural network.According to the characteristics of pollutant variation and existing theory,we propose a method which uses the 24-hour pollutant concentrations of the prior day to predict the present day.Based on the Nanjing pollutant concentration data,we have a detailed discussion on the selection of structure and parameter,and make some improvements on the BP network.The model runs well on Nanjing pollutant concentration data.2)To improve the prediction accuracy,we propose an air pollutant classification method based on Kohonen neural network,which classifies the original data.We also propose a method adding meteorological factors to improve the prediction.Using PCA method we extract three principal components and add to the original BP neural network.Final results show these methods improve the prediction accuracy and the meteorological factors improve generalization.3)The study sets in the National 863 Project background,running on the"Environment Cloud".We analyze the technology architecture of the"Environment Cloud".We demonstrate the running results on the platform.The successful application indicates usability,versatility and effectiveness of the model.Mutually it proves that the "environment cloud" brings innovation and change to the environmental field.