Research On Prediction Of PM2.5 Multivariate Chaotic Time Series Based On Phase Space Refactoring

In recent years,there's frequent fog and haze in some provinces and cities of China.The real-time detection and release of PM2.5 concentration is increasingly unable to meet people's travel needs,and a timely and effective PM2.5 concentration prediction mechanism and method is needed.Since PM2.5 is affected by multiple and complex factors such as meteorological elements and other pollutants in the air,and those factors are with the complex and non-linear characteristics between each other.The production of PM2.5 involves a large number of nonlinear chemical and photochemical kinetics,which is with chaos,so it's extremely difficult to predict PM2.5concentration.In order to more effectively analyze change trend of PM2.5 status for a time in the future and predict PM2.5 concentration,the author in this dissertation introduced the chaos theory and the prediction method of chaotic time series to establish a prediction model of the multivariate time series for PM2.5 concentration,and took the prediction of PM2.5 concentration in Beijing as an example to a-chieve its short-term forecast.An idea of prediction model construction for the multidimensional factors was proposed according to the complexity of the air system.Considering that both PM2.5 and its affecting factors were considered as comprehensively as possible,the phase space law of the air system was explored based on the idea of chaos prediction,and the prediction model was established on this basis.First,in order to fully explore the evolution trajectory of the air system,the phase space of the unit-related chaotic time series was extended to the phase space of the multivariate time series,and the phase space matrix of a multivariate time series was constructed,so that the system chaotic attractors can be recovered effectively.Second,the RBF neural network was used to predict the state points in multivariate phase space,the state point of the previous phase space was used as neural network input,the state point of the next phase space was considered as the neural network output to train the network.The trained network was used to predict the phase space point of the sample,and finally the phase space point of the time series to be predicted was separated to realize the prediction.Third,the constructed prediction model was used in the classic Lorenz chaotic system to initially verify the reliability of the model.Secondly,by taking Beijing as an example,this data on air pollutants and historical meteorological hours in Beijing from May 2014 to November 2014 was applied in this dissertation to address the missing data values,and key factors affecting the PM2.5 concentration in Beijing were analyzed,and the correlation between PM2.5 concentration and the groups of affecting factors were analyzed,together with the judgment of their chaotic characteristics,and the eigenvector of the prediction model input was ultimately determined,the prediction model of multivariate chaotic time series constructed in this dissertation was used to predict hourly PM2.5 concentration in Beijing.Finally,in order to verify the reliability of prediction model of PM2.5 concentration proposed in this dissertation,the unit prediction method of chaotic time series,prediction model of multivariate chaotic time series and the commonly used ARIMA model of unit time series,prediction model of multivariable linear regression time series were used as the comparison of the empirical prediction experiments.The precision of prediction model was measured based on mean-square error(MSE)and mean absolute deviation,and accurate evaluation was given to different prediction models.The results show that the root mean square error of the multivariate chaotic time series prediction model based on phase space reconstruction in PM2.5 concentration prediction is 4.92,and the average absolute error is 2.4,which is more accurate than the traditional statistical forecasting method.