Research On Air Pollution Forecasting Based On Artificial Neural Networks

With the development of society and economy and the population expansion of the city, air pollution becomes more and more serious, therefore, it is significant to study air pollution forecasting so as to understand environmental variation trend, prevent and control air pollution. Generally the atmospheric environmental system is complex, and now mass historical monitoring data has accumulated, therefore, the traditional forecasting methods have difficulty in excavating useful information from the mass data to forecasting precisely. This paper applied the artificial neural network (ANN) in the air pollution forecasting, presented an air pollution forecasting method based on artificial neural networks in virtue of the ANN's good nonlinear processing ability and fault-tolerant ability, and then provided a accurate and comprehensive environmental quality information for making decision.In this thesis, forecasting model was built up using artificial neural networks to forecast the hourly concentration of PM2.5. The model was fixed based on the relationship between the hourly concentration of PM2.5 at any hour of a day and the 24 hourly concentrations measured on its previous day. The data of hourly concentrations were obtained from monitoring sites of London Air Quality Network. Firstly, Error Back Propagation network (BP network) was selected as pollutant concentration forecasting network. The methods of setting the network's structure were confirmed and a strategy of training BP networks using Bayesian Regularization method and Early Stopping method were presented based on large numbers of experiment. Moreover, how to initialize the weight, divide the sample data and divide it in proper ratio were discussed. All these studies provide important experiences for later study on model and network structure parameters selection. This paper also put forward a method of dividing a year into two parts using self-organization competitive network and build forecasting model according to seasons respectively. It can improve the network's forecasting performance obviously.Secondly, the impacts of numbers of sample data, noise-reduction and weather factors on the networks forecasting performance were discussed respectively. The computation results show that the forecasting networks of PM2.5 have good quality on forecasting precision and generalization ability.Finally, according to the previous methods for model building we presented, forecasting networks were built up to forecast the hourly concentrations of PM10, NOx and O3.This research results show that it is reasonable to apply the artificial neural network in air pollution forecasting. It provides a new method for urban air pollution forecasting in information society, and a feasible method to make the most use of mass environmental data.