Research On PM_2.5 Concentration Prediction Using Multi-source Data And Ensemble Self-organizing Fuzzy Neural Network

In recent years,air pollution is frequently appeared in our daily lives,which directly affects people’s physical health and normal activities.As an important factor causing air pollution,PM_2.5 has gradually become the focus of people’s attention.The prediction results of PM_2.5 concentration cannot only effectively guide residents to reasonably arrange daily activities,but also provide an important reference for the government departments to take effective prevention and control measures in time.Therefore,the prediction of PM_2.5 concentration has important practical significance and social value.1.Since the PM_2.5 concentration is associated with many interconnected complex variables,a hybrid feature selection algorithm based on scatter plot（SP）and partial mutual information（PMI）is designed to select the critical variables.Firstly,the correlation between the concentration of PM_2.5 and each variable in the original multi-source feature set is qualitatively analyzed by the scatter plot,and then the preselected feature subset is selected from the original multi-source feature set.Secondly,the partial mutual information method is used to quantitatively calculate the partial mutual information value between the concentration of PM_2.5 and each variable in the preselected feature subset.Moreover,the variables are filtered based on the changes of T_AIC values to determine the optimal feature subset.Finally,some other methods are introduced and compared with the proposed feature selection method.The experimental results show that our method can effectively perform feature selection comparing with other methods.2.A self-organizing fuzzy neural network base on hybrid evaluation index（HEI-SOFNN）is proposed.Firstly,the network parameters are initialized by the fuzzy C-means clustering（FCM）in order to solve the issues,such as trapping into local minima and slow convergence,caused by inappropriate initial parameters selection.Then a novel relevance evaluation index（REI）is presented to calculate the correlation among the outputs of neurons in the RBF layer.The learning ability of neural network will be determined by the change of root mean square error（RMSE）during the training process.Furthermore,a hybrid evaluation index（HEI）is proposed based on REI and RMSE.The topology structure of the fuzzy neural network is adjusted according to the HEI.In order to ensure the convergence speed and accuracy of the neural network,an adaptive gradient descent algorithm is designed to update the nonlinear parameters of the network and the least square method is used to update the linear parameters of the network.In addition,the convergence analysis of the presented HEI-SOFNN is also provided to ensure the performance and the reliability of the network.Finally,three benchmark experiments are used to verify the effectiveness of the designed network.3.An ensemble self-organizing fuzzy neural network based on the Bagging ensemble mechanism and self-organizing fuzzy neural network（EHEI-SOFNN）is proposed.Firstly,the bootstrap sample method is used to obtain some diverse sample subsets,which can train several weak learners with differences.Secondly,the Bagging ensemble mechanism and HEI-SOFNN are introduced to design the EHEI-SOFNN.At the same time,the convergence proof of the EHEI-SOFNN network is given to guarantee the performance and reliability of the network.Finally,three benchmark experiments are used to verify the effectiveness of the designed network.4.The prediction model of PM_2.5concentration based on EHEI-SOFNN is established.Firstly,taking Beijing city as the research area,the hourly multi-source data,including image data,meteorological data and pollutant concentration data,are collected.Moreover,the hybrid feature selection method is used to obtain an optimal feature subset with 12-dimensional features.Secondly,the optimal feature subset is used as input of the EHEI-SOFNN to predict the hourly concentration of PM_2.5.Finally,the results of experiments show that the proposed prediction model of PM_2.5concentration is effective and can accurately predict the hourly concentration of PM_2.5comparing with other algorithms.5.A PM_2.5 intelligent prediction APP is developed.Firstly,the demand analysis of the PM_2.5 intelligent prediction APP is given and its main function is to provide the users with the prediction results of PM_2.5 concentration quickly and easily.Secondly,the development plan and the technical route of APP are designed and improved.Finally,the parameters of EHEI-SOFNN model trained by MATLAB are stored in the Apache server.The Android language is used for client development in the Android Studio development environment.Then,an intelligent prediction app is launched,which has the functions of predicting the hourly concentration of PM_2.5,querying weather conditions and air quality,and pushing life suggestions.