Research On The Data-Driven Modeling And Prediction Of COVID-19 Epidemic

An unknown Novel Coronavirus(SARS-Co V-2)was first reported in Wuhan,Hubei Province at the end of December,2019,and then spread rapidly.As of March 20,2022,the COVID-19 pandemic has caused over 400,000 confirmed cases and claimed more than 10,000 lives in total in China.From a global perspective,there have been over 400 million confirmed cases and over 6 million cumulative deaths.At present,the pandemic situation is still grim.The accurate modeling and prediction of the evolution law of COVID-19 can provide important references for the scientific formulation of policies and measures concerning pandemic prevention and control as well as long-term/short-term treatment plans.Therefore,it is of great significance to study the evolution law of COVID-19 transmission and design infectious disease models with higher prediction accuracy.From the data-driven perspective,the transmission law of the COVID-19 pandemic was modeled and studied based on the traditional dynamics theory of infectious diseases and the deep learning theory in this paper.The main research contents are summarized as follows:(1)In terms of the traditional dynamic model,an improved SEIR model was constructed based on symptom classification in this paper.Firstly,combined with the reality of symptom classification and graded treatment of patients with COVID-19,an improved SEIR model was established by introducing three compartments corresponding to mild,severe and critical illness,respectively,into the traditional SEIR model.Then,the Runge-Kutta method was adopted to solve the differential equation model,and the nonlinear least square algorithm was designed to estimate the model parameters according to the official data published by the National Health and Health Commission of the People’s Republic of China.Finally,the COVID-19 pandemic was modeled and predicted based on the proposed SEIR model.The experimental results revealed that the introduction of symptom classification compartments into the SEIR model can not only effectively improve the prediction accuracy but also accurately describe the evolution law and mutual transfer law of patients with different symptoms.This model can provide more accurate and comprehensive theoretical support for the decision-making of government departments,especially for the construction of mobile cabin hospitals and the rational preparation of important pandemic prevention resources in the case of a large-scale outbreak of the COVID-19 pandemic.(2)In terms of the deep learning model,the Bayesian Weighted-LSTM(BWLSTM)multi-sequence COVID-19 pandemic prediction model was constructed in this paper.Firstly,the BW-LSTM multi-sequence COVID-19 pandemic prediction model was established by introducing weight coefficients for each output dimension.Then,a super-parameter optimization learning algorithm of the LSTM network was designed based on the Bayesian optimization method to learn the model super-parameters including dimension weight coefficients.After that,the optimal network-parameter learning of the LSTM multi-sequence weighted prediction model was realized through reasonably setting the objective function of Bayesian optimization.Finally,the modeling and prediction of the COVID-19 pandemic were realized based on the proposed BW-LSTM model.The experimental results showed that on one hand,the introduction of dimension weight coefficients can effectively reduce the adverse effects of complex correlation between sequences on modeling results;on the other hand,the introduction of Bayesian optimization algorithm can realize efficient learning of network hyper-parameters and effectively avoid over-fitting.(3)In terms of the parameter identification of the epidemic evolution model,this paper proposed a parameter identification model based on LSTM by using its powerful data learning and knowledge transfer ability.Firstly,the input and output of the LSTM network were designed reasonably to make the LSTM network have the ability of parameter identification.Then,the sample data set with good representativeness and completeness was obtained based on the SEIR model simulation.After that,the LSTM parameter optimization algorithm was realized through Bayesian optimization so as to obtain the final LSTM parameter identification model.Finally,the validity of the model was analyzed and verified based on artificial data and real-life data sets.The results indicated that the proposed parameter identification model based on LSTM can utilize the rules learned from artificial data to realize parameter identification in the case of few sample data,with the identification accuracy significantly better than that of the traditional least square method.