Prediction Of COVID-19 Spread Based On Improved SEIR Model And LSTM Model

The COVID-19 pneumonia is spreading worldwide,which is a communicable disease that spread rapidly and endanger human health.Until 2021,the COVID-19 pneumonia epidemic was spreading rapidly in many countries and territories.Therefore,the trend of development and spreading of the COVID-19 pandemic is urgently needed to be studied.In addition,we need to develop a long-term prediction model to predict the number of diagnosed cases in the future,predict when the inflection point of the COVID-19 pandemic may occur,and analyze the impact of a series of prevention and control measures introduced by the government on the prevention and control of the COVID-19 pandemic,which also has a significant impact on the national economy.In order to explore whether these measures have had an impact and how long they should last,it is very important to analyze and predict the development trend of the COVID-19 pandemic.The data sources about the COVID-19 pandemic in this paper were epidemiological data provided by China Health Commission and the Johns Hopkins University.At present,most of the COVID-19 prediction models only used cumulative number of diagnosed cases to train,which led to the training model can only predict the rising trend of the epidemic,but cannot predict when the epidemic may have inflection point.Therefore,the third chapter of this article uses the SEIRD model to predict the development trend of the epidemic and the possible inflection point.The fourth chapter of COVID-19 pneumonia transmission trend prediction model in this study was established based on the long and short term memory network（LSTM）with the daily number of new diagnosed cases training set and the Google Trend of COVID-19 as the keyword.Secondly,most of the prediction models were based on the historical cumulative number of diagnosed patients to train the models without considering the real-time update of the input sequence.However,the prediction of the current time is likely to have a certain impact on the accuracy of the prediction of the next time,so the rolling update mechanism was adopted in the study,the current prediction results were used to update the input sequence in real time to make long-term accurate prediction.Thirdly,considering that the historical data of different time points have different contributions to the data of the current prediction time points,the attention mechanism was introduced.The input influence weight was given to highlight the effective features and improve the prediction accuracy by employing the attention mechanism.Fourth,the SEIRD model was used to change the degree of intervention by controlling the exposure rate to reflect the significance of prevention and control.The SEIRD model（Susceptible-Exposed-Infected-Recovery-Death model）designed in this paper is based on the basic SEIR model（Susceptible-Exposed-Infected-Removed model）,but the Remover group in the basic SEIR model is subdivided into Recovery group and Death group,and ues the simulated annealing algorithm and the basic regeneration number R₀for parameter estimation,and the prediction error rate of the SEIRD model proposed in this study is 6.67%in predicting the COVID-19 epidemic situation in Hubei Province,however,the prediction error rate of the SEIR model embedded in the number of immigrants and emigrants proposed by Yang Zifeng et al.is 12.56%.In predicting the COVID-19 population in Iran and Peru,the average absolute error rate of RA-LSTM（Rolling-Update Attention Long Short-Term Memory）model proposed by this study is 2.45%,while the average absolute error rate of the improved LSTM model proposed by Wang Peipei et al.is 4.69%.The performance of RA-LSTM proposed in this paper has advantages.