Nonlinear Propagation Dynamics Analysis And Prediction Of COVID-19 Based On Complex Network

Since the outbreak of the COVID-19 in December 2019,the COVID-19 epidemic has spread rapidly to different countries in the world.The cumulative number of confirmed cases in the world has reached more than 400 million,and the cumulative deaths toll has reached more than 5 million,which has caused great pressure on the global public health system.In addition,to suppress the spread of COVID-19,corresponding measures and policies are adopted by differet countries’s governments.These measures have caused the global economy to stagnate and trade to be interrupted,causing huge losses to the global economy.Therefore,study the internal transmission mechanism and infection characteristics of COVID-19,build a new prediction method according base on its transmission mechanism,infection characteristics and historical case data,that will accurately predict the trend of the epidemic in the future.This is of great practical significance for government departments to formulate scientific and effective policies for epidemic prevention and control.In this paper,based on nonlinear Granger causality analysis,detrended cross-correlation analysis(DCCA),complex network analysis and LSTM model,a novel mixed prediction model(Network-LSTM model)is constructed.The model is used to predict the number of confirmed cases of COVID-19 in 51 States(including one special administrative region)in the United States.The major work is as below:First of all,descriptive statistical methods are used to analyze the cases of COVID-19 epidemic infection in 51 states of the United States from January 21,2020to August 11,2021.The results showed that the sequence of infection cases in each state does not obey the typical random normal distribution,but presents complex nonlinear characteristics.Secondly,based on the nonlinear Granger causality test and DCCA method,the nonlinear spread network of COVID-19 in the United States is constructed,and its spread characteristics are researched.The results show that the nonlinear spread network of COVID-19 in the United States has typical small-world characteristics.In addition,it is found that California,Colorado and New York are playing important roles in the spread of the epidemic in the United States.Finally,through the comparative analysis between the aviation network and nonlinear spread network of COVID-19 in the United States,it is found that long-distance average activities in human behavior dynamics may be the main dynamic mechanism leading to the evolution of COVID-19epidemic in the United States.Thirdly,Based on the sliding window method,the time-varying of COVID-19nonlinear spread network is analysed.The results showed that the evolution trend of network structure index showed good consistency,and its trend is very similar to the actual epidemic development process.The results show that the complex network evolution index can effectively reflect the dynamic evolution characteristics of the mutual influence of the COVID-19 epidemic among American states.Therefore,based on nonlinear spread network of the COVID-19,a feature extraction layer can be established to extract the evolution features of COVID-19 epidemic spread interaction between states.Fourth,the prediction model based on the mixed features of COVID-19 and LSTM Model,a novel Network-LSTM hybrid prediction model with higher accuracy is constructed.The Network-LSTM model and traditional LSTM is used to predict and analyze COVID-19 epidemic infection cases in American States from January 21,2020to August 11,2021,and through mean absolute error(MAE),root mean square error(MAPE),correct rate(%)and coefficient of determination(R~2).This four indicators comprehensively evaluate the advantages and disadvantages of the two kind of prediction models.The results show that compared with the LSTM model,the MAE,MAPE,andR~2 of the Network-LSTM hybrid model have been improved by 18.51%,19.97%,5%and 0.02 respectively.This means that the prediction performance of the Network-LSTM hybrid model is improved significantly.It also shows that the embedding of the feature extraction layer can significantly improve the prediction performance of the prediction model.