Study On Long And Short-term Memory Model Based On Soft Sharing Mechanism And Its Application In Predicting COVID-19

The Corona Virus Disease(COVID-19)continues to spread asynchronously around the world,causing economic downturns and trade growth decelerating and the probability of chain enterprise failures increasing greatly.Predicting the trend of COVID-19 through system modeling can effectively simulate the development,predict the spread,and assist in the control of the epidemic.In the application of current common methods to predict COVID-19: infectious disease dynamics methods rely excessively on accurate estimation of dynamic parameters;traditional statistical methods are difficult to meet the time series stationarity assumptions;and machine learning methods require a lot of data although they have relatively few limitations.In addition,most of the existing works focus on the fields of cumulative confirmed cases and single-step prediction,while few are involved in the new confirmed cases and multi-step.This paper studies the multi-step prediction methods of Long and Short-Term Memory network(LSTM)and its application in the trend prediction of COVID-19 new confirmed cases,and proposes the Multi-Task Long and Short-Term Memory network based on Soft Sharing mechanism(SS-LSTM),which can solve the problem of insufficient data of cowid-19 to a certain extent.The main content and contributions of this paper are as follows:(1)This paper studies methods of Recursive Prediction LSTM(RP-LSTM),Vector Output LSTM(VO-LSTM)and Encoder-Decoder LSTM(ED-LSTM),and its application in the new confirmed COVID-19 case using datasets of Brazil,Chile,India,the United Kingdom and the United States.The optimal hyperparameters were explored through Bayesian Optimization.Experiments show that,in 5 countries,the ED-LSTM has the best overall performance(MAPE is 15.74%,9.73%,13.01%,16.29%,19.26%),the VO-LSTM is second(14.44%,20.24%,14.78%,25.07%,32.86%),the RP-LSTM is the worst(33.80%,21.71%,33.08%,37.88%,31.24%).In addition,the cumulative error of ED-LSTM is lower than others,the error of the multi-step prediction process is more stable,which means the ED-LSTM is more suitable for such task.(2)In view of the insufficient amount of data in the application of the LSTM multi-step methods to predicting COVID-19,the SS-LSTM network is proposed for implicitly increasing the amount of data and improving the performance.Experiments show that the proposed network not only has higher accuracy(MAPE is 7.87%%,13.27%,7.43%,9.23%,16.17%),and can further improve the problems of cumulative error in the multi-step prediction of COVID-19.(3)The proposed SS-LSTM model has certain reference value in the comprehensive study of COVID-19 transmission trend prediction in different regions.