Research On PM2.5 Concentration Prediction Using A Self-organizing Recurrent Fuzzy Neural Network

In recent years,the fog and haze weather has appeared in our life with higher and higher frequency and larger range,which seriously affects health and normal production and life.And air quality has gradually become a hot issue that people are concerned about.As the main factor causing haze weather,the concentration of PM2.5 is influenced by many factors such as direct pollutant discharge,secondary transformation of pollutants,meteorological conditions,topography and terrain.It has obvious nonlinear dynamic characteristics and is more difficult to predict.Therefore,this paper studies the prediction method of PM2.5,and completes the development of the prediction software for applying the research results to our daily life.1.The filter-wrapper mixed feature selection algorithm based on mutual information and particle swarm optimization(PSO)is proposed in response to the complex problems of the influence variables of PM2.5.In the filter phase,the maximum correlation of each variable for PM2.5 and the redundancy between variables are taken into full consideration.Based on the maximum correlation minimum redundancy(MRMR)principle of mutual information,the variables are sorted,and the primary selection of the feature subset is carried out.In the wrapper stage,the initial selected feature subset is used as the initial search range,and the predictive precision of neural network is used as the fitness function of the PSO optimization algorithm to determine the optimal specific subset.Finally,the feature selection method is compared with other methods,and the effectiveness of the method is verified.2.A self-organizing recurrent fuzzy neural network based on the completeness of fuzzy rules and partial least squares(PLS)is designed.Firstly,in order to enhance the dynamic information processing capability of the network,a recurrent fuzzy neural network(RFNN)is composed of feedback connections with internal variables in the rule layer of the basic fuzzy neural network.Then,aiming at the problem that the structure of RFNN is difficult to determine,a self-organizing mechanism based on the completeness of fuzzy rules and PLS is proposed.At the same time,in order to accelerate the convergence speed of neural network in the training stage,the gradient descent algorithm with adaptive learning rate is used to adjust the network parameters.Then,through two classical nonlinear experiments,the effectiveness of the designed neural network is verified in terms of network structure and approximation accuracy.3.The prediction model of PM2.5 concentration is established.Taking an air quality monitoring station in Beijing as the research object,the hourly meteorological data and pollutant data of the station are collected,and the characteristics selection method of filter-wrapper is used.Using the prediction accuracy of self-organizing recurrent fuzzy neural network(SORFNN)as the fitness function of PSO,then the 7-dimensional feature subset is determined,and the 7-dimensional variables are used as the inputs of the network to predict the PM2.5 hourly concentration.Compared with other algorithms,the proposed prediction model can effectively predict the hourly concentration of PM2.5.4.A smartphone application software for PM2.5 intelligent prediction is developed,which enables users to know the PM2.5 prediction concentration more conveniently and quickly.The software is based on the SORFNN prediction method and developed with java language in the development environment of Eclipse Java EE.It includes three modules,model training,server and android client.And it can predict the PM2.5 concentration,query weather and air condition in real-time,query historical data and so on.