Prediction Of Urban Air Pollution Concentration Based On Neural Network Model

Air quality is widely concerned in daily life.Air pollution can not only seriously affect people’s physical and mental health,but also cause serious harm to plants and animals.With the development of new information technologies such as big data and deep learning,how to use air pollution big data and deep learning technology to make scientific and effective prediction of urban air pollutant concentration is a hot issue in the field of air pollution control.This paper mainly studies the characteristics of spatiotemporal correlation and nonlinearity of air pollution data,and selects Long Short-Term Memory（LSTM）as the basic model to make real-time prediction of multivariate time series data.Make full use of the long-term memory ability of the long-term and short-term memory network to carry out the representation learning of the input data,and tap the potential change rules of the data.This paper mainly studies the following aspects:（1）Collected real-time historical data of urban air pollution and meteorological data from the China National Environmental Monitoring Station and National Climate Data Center（NCDC）,learned about the relevant theoretical knowledge and technical principles of air pollutant concentration prediction,the Pearson correlation analysis is performed on the characteristic factors that affect the diffusion of air pollutant,which provides a theoretical basis for the feature selection of the prediction model.（2）An LSTM urban air pollutant concentration prediction model based on time series decomposition is proposed.First of all,Based on the long short-term memory neural network,the Complementary Ensemble Empirical Mode Decomposition（CEEMD）is introduced,it is used to smooth the air pollutant concentration data.Then,the construction process of CEEMD＿LSTM prediction model and related parameter settings are described in detail.The concept of"decomposition and integration"is used to predict PM_2.5 and other concentrations.The experimental results show that the CEEMD＿LSTM prediction model is applicable to the prediction of air pollutant concentrations such as PM_2.5,PM₁₀,SO₂,etc.Finally,in order to further demonstrate the superiority of the prediction model,the CEEMD＿LSTM prediction model is compared with other neural network models（BP neural network,RNN neural network）and the traditional machine learning model SVR,etc.,and the comparison shows that the CEEMD＿LSTM model has the best prediction effect.（3）From the perspective of input data,an optimization strategy based on time,space is established for the model,and a CEEMD＿LSTM prediction model based on space-time optimization is proposed.By exploring the influence of different sliding time window T on the model,the model is optimized based on time,and the air pollution data and meteorological data of surrounding cities are introduced for spatial optimization.The experimental results show that the CEEMD＿LSTM prediction model based on space-time optimization has the best performance,followed by the time and space-based prediction model.（4）The Stack Auto Encoder neural network is introduced to construct a CEEMD＿LSTM prediction model based on feature reduction.Firstly,pre-trained SAE is used to perform data dimensionality reduction,then the structure of the SAE＿CEEMD＿LSTM model is constructed,and finally the PM_2.5 concentration is predicted through experiments.The experimental results show that the improved SAE＿CEEMD＿LSTM model has higher prediction accuracy.