Research On Air Pollution Feature Classification And Deep Learning Prediction Based On Candlestick Chart And Mean Line Theory

With the development of China’s economy,air pollution is also emerging.As one of the air pollutants,PM_2.5 concentration in the air directly determines the degree of air pollution.Long-term PM_2.5pollution weather will seriously affect human health.In recent years,PM_2.5 pollution has also appeared in some tourism cities with low degree of industrialization in southern China.Therefore,in order to grasp the change characteristics of PM_2.5and effectively predict the change of PM_2.5concentration,this paper uses the PM_2.5data of Guilin to carry out a series of studies.The main research contents are as follows:（1）Research on the change law of PM_2.5 under different time scales.This paper has compiled the PM_2.5concentration data of 44 automatic monitoring stations in Guilin urban area from 2013 to 2018,and studied the annual change characteristics of the PM_2.5concentration changes in Guilin urban area during the six years.And selected PM_2.5concentration data from December 1,2017 to November 30,2018 as basic research data,and analyzed the seasonal and daily characteristics of PM_2.5concentration.Finally proved that PM_2.5 data has a certain periodic change characteristic.（2）Study on the influencing factors of PM_2.5change.In this paper,the daily pollutant concentration data and meteorological data at the monitoring station of the Eighth Central Committee of Guilin from December 2017 to November 2018 were used as the baseline data to analyze the correlation between PM_2.5concentration and them.The results showed that PM_2.5 concentration was positively correlated with SO₂,NO₂,PM₁₀ and CO concentrations,but not significantly correlated with O₃and O₃＿8 H concentrations.At the same time,PM_2.5concentration was significantly correlated with wind speed,wind direction,temperature,humidity and rainfall,but not significantly correlated with air pressure.（3）Research on feature extraction method of PM_2.5 change.In this paper,two methods are used to extract the variation characteristics of PM_2.5:one is to extract the time domain and frequency domain characteristics of PM_2.5 at the time-frequency analysis level.The mean line and waveform theory are introduced to extract the time domain characteristics of air pollution data from geometric aspects,and the discrete Fourier transform is used to extract the frequency domain characteristics at the frequency domain analysis level.Two is to construct a Candlestick chart generator to extract the Candlestick chart features of PM_2.5data.The Gaussian diffusion model is established to simulate the diffusion process of PM_2.5,and the basic form of K line is explained by the diffusion process of different PM_2.5.A total of 6 time-domain features,3 frequency-domain features and 16 Candlestick chart combinations that can reflect the physical transmission law of PM_2.5are extracted as the Candlestick chart features of PM_2.5 data.The variation trend of PM_2.5 concentration in the next three days can be accurately predicted according to these 16 combination characteristics.Using the average concentration of PM_2.5in the next three days as an evaluation index,the accurate data for predicting the future trend reached 99.68%.（4）Research on PM_2.5 concentration prediction.In this paper,the PM_2.5 concentration data,the concentration data of other atmospheric pollutants and meteorological factors screened,the time-domain characteristics,the frequency-domain characteristics and the K-line characteristics are used as the input variables to construct the VGG-LSTM combination model to classify the extracted characteristics and predict the future PM_2.5 concentration.At the same time,artificial neural network（ANN）,recurrent neural network（RNN）and long short-term memory network（LSTM）models were constructed for comparative experiments.The scatter plots of the real and predicted PM_2.5 values of the four models were also compared.The experimental results show that the average absolute error（MAE）of VGG-LSTM model is 10.5971,the root mean square error（RMSE）is 12.561,and the symmetric average absolute percentage error（SMAPE）is 0.2024.Among them,compared with the three models of ANN,RNN and LSTM,the VGG-LSTM model has an average increase of20.89%in MAE,20.46%in RMSE and 31.25%in MAPE.This proves that the VGG-LSTM combination model has good prediction performance.In summary,this article has completed the extraction of atmospheric pollution characteristics by studying the periodic characteristics,time domain characteristics,frequency domain characteristics and K-line characteristics of air pollution.In the research process,the geometric meaning of the waveform was abstracted into the feature extraction method of the air pollution data waveform by using the mean line and waveform theory,and the K-line diagram that can reflect the transmission process of the air pollution was used to extract the features of the air pollution,establishing a practical A feature set of high-degree atmospheric pollution.By using these feature sets,a VGG-LSTM combined model is constructed to classify and predict air pollution,achieving the purpose of precise prediction of short-term PM_2.5 concentration.