Research Of PM_2.5 Prediction Based On Second-order Self-organizing Fuzzy Neural Network

China has suffered the problem of conspicuously rising air pollution.The regional atmospheric environmental problems caused by PM_2.5 have been increasingly prominent.Chronic exposure to PM_2.5 is leading serious health issues to the public.PM_2.5 also affects climate and limits the sustainable development of society and economy.In order to take preventive and regulatory measures to manage air quality,it is highly required to develop methods that can predict PM_2.5concentrations precisely.However,estimating the PM_2.5 distribution has proven to be tough due to the fact that the PM_2.5 concentrations are simultaneously affected by emission sources,air pollutants,meteorological conditions and topographical characteristics of the area under study.Thus,a self-organizing fuzzy neural network with second-order gradient?SOG-SOFNN?algorithm based PM_2.5 soft senor model is desired to predict the hourly concentrations of PM_2.5 for the next 24 hours.This thesis further designs the PM_2.5 intelligent prediction visualized system.The main research contents of this thesis are included as follows:1.Feature extraction of the PM2.5 soft sensor model.The site selected for applying the proposed methodology is called Shangdianzi,which is a regional background station.The sampling time is from 14 January to 23 January 2010.Firstly,the relationship between PM_2.5 concentrations and meteorological variables,the components of PM_2.5 and its precursor gas and the interaction between PM_2.5concentrations and aerosol optical depth were analyzed.Then,the hourly monitoring data of meteorological variables and air pollutants as well as aerosol optical depth were collected at Shangdianzi during the studied period.In order to eliminate the irrelevant input parameters and tedious information,the principal component analysis?PCA?is used to select the dominating variables most correlated to PM_2.5 as the input variables of the PM_2.5 soft sensor model.It could be found that the dominating variables?relative humidity,pressure,aerosol optical depth,wind speed,wind direction?extracted by PCA have a high correlation with the characteristics of PM_2.5 at Shangdianzi where the PM_2.5 concentrations are severely correlated to the meteorological parameters and the aerosol optical depth.2.Frame construction of the SOG-SOFNN.The SOG-SOFNN with adaptive structure is used to establish the PM_2.5 soft sensor model to understand the dynamic process that produces PM_2.5.The SOG-SOFNN splits or deletes neurons in normalized layer according to the contribution to the network output made by outputs of neurons in this layer,which is determined by the sensitivity analysis of model output.The parameters of the SOG-SOFNN is optimized by a second order gradient descent algorithm with improved training speed and higher approximation accuracy.The experimental results on nonlinear dynamic system identification and Mackey-Glass chaotic time series prediction show that the proposed SOG-SOFNN performs better than other self-organizing fuzzy neural network?SOFNN?with more compact structure and higher prediction accuracy.3.Soft sensor modeling of PM2.5 based on SOG-SOFNN.Due to the predicting of PM_2.5 in the atmosphere is tough,the SOG-SOFNN is developed to forecast the hourly PM_2.5 concentrations of the next 24 hours using the current values of dominating variables at Shangdianzi during the studied period.The experimental results of the PM_2.5 soft sensor model applied to measured data show that the method can forecast PM_2.5 in real time.Compared to other neural networks?NN?,the SOG-SOFNN based PM_2.5 soft sensor model has better training effect and higher prediction precision.It can be concluded that the SOG-SOFNN presented here is valid for estimating the hourly PM_2.5 concentration.4.Visual design of the PM2.5 intelligent prediction system.The SOG-SOFNN based PM_2.5 soft sensor model is applied to develop the PM_2.5 intelligent predicting system to visualize the PM_2.5 prediction results.Firstly,build Tomcat server to store the test data of the PM_2.5 soft sensor model.Then,use Eclipse Java EE to compile Android client which embeds the well trained PM_2.5 soft sensor model.The Android client acquire the test data from the Tomcat server through HTTP request to make prediction of PM_2.5 concentrations.Finally,users can possess more convenient and direct assess to get air quality information through the predicted value of PM_2.5 and air pollution condition as well as the travel advice provided by the PM_2.5 intelligent predicting system.