Cooperation or confrontation? —— Prediction of COVID-19 pandemic situation via deep learning

The COVID-19 pandemic is the most serious catastrophe since the Second World War. To more accurately observe the epidemic under the influence of policies and provide policy adjustments before the official presidential transition in the United States, we use a three-layer superimposed Long-Short-Term-Memory (LSTM) model to predict the epidemic development trend to mid-January, 2021. The proposed model provides more accuracy and stability relative to Susceptible-ExposedInfective-Recovered (SEIR), modified stacked au-to-encoder, and single-layer LSTM models. The performance effects of the measures in China and five countries with severe epidemics are analysed and summarised. The model shows that the error rate of China, five countries and the world is less than 1.4%. According to forecasts, the epidemic situations in the United States, India, and Brazil, caused by untimely, inappropriate policies, lax regulations and insufficient public cooperation, remain very severe, with cases continuing to increase by tens of thousands. The number of cumulative confirmed cases worldwide will exceed 84.58 million by mid-January, 2021; however, the mortality rate will gradually decrease. Based on analysis of measures (including China’s effective prevention and control policies), we found that there are performed tremendous different efficiency even using same positive policy for different countries because of various cooperation between people and governments. It is essential to maintain self-protection to prevent the epidemic from deterioration or regenerating, especially, wearing mask and maintaining a safe distance. Introduction The coronavirus disease 2019 (COVID-19) is brought on by infection from Severe Acute Respiratory Syndrome (SARS) Coronavirus 2, and the reports of related cases were first released by Wuhan, Hubei Province, China in December 2019. The widespread COVID-19 epidemic is a serious threat, and has become one of the most challenging global catastrophes facing mankind since the Second World War. On March 11, 2020, the global development of COVID-19 was assessed by the World Health Organization (WHO) as having met the characteristics of a pandemic. The mortality rate of COVID-19 is estimated to be between 2% and 5%, i.e. lower than that of SARS and Middle East Respirator Syndrome . However, COVID-19 has a higher infection rate than bat-like SARS, and its pathogenicity is between that of SARS and bat-like SARS. Patients with COVID-19 generally have symptoms such as fever, dry cough, fatigue and dyspnea, the immunological function is weakened, and in severe cases, respiratory failure and death may occur. The heart may be damaged and develop myocardial inflammation after recovery from COVID-19. In addition, the slowdown or stagnation of national economies during the epidemic prevention period has caused a deceleration in global economic growth, along with aggravating unemployment and hunger. Some national leaders have been infected with the virus, thereby affecting the implementation of policies and management. On January 15, 2020, the Chinese Center for Disease Control and Prevention (China CDC) initiated a first-level emergency response. China began implementing measures since January 23, 2020, and the epidemic had stabilized by March, 2020. In contrast, the current epidemic situation overseas is not very optimistic. The number of infected people continues to maintain substantial increase of tens of thousands. According to the report by the WHO on December 6, more than 200 countries, territory and regions have been invaded by COVID-19, the global epidemic has infected more than 65.8 million people, with 1523656 deaths, and the epidemic in the Europe, Americas and South-East Asia are the worst, especially in the United States and India. At present, 52 candidate vaccines have entered phase clinical trials; 13 of these have entered the third phase. It is expected that vaccines may be applied by the end of 2020. In addition, countries with outbreaks have taken many measures, such as city lockdowns, banning unimportant gatherings, and business opening in advance, but the effect of the policies are unclear. The president of the United States will officially transition on January 20, 2021, and new policies affecting the epidemic situation may be promulgated. Therefore, the aim of this study is to predict subsequent trend according to the epidemiological data of COVID-19, and to conduct a policy analysis to summarise meaningful and positive measures. This information has great reference significance for assessing the development of the epidemic situation in China, and in five countries with serious epidemic situations. According to the forecast, it will be helpful to assess the effectiveness of current policies and make flexible adjustments, for instance, reinforcement, alteration, and relaxation of restrictions. According to incomplete statistics, there are three main methods for forecasting the epidemic development of COVID19, as follows: • The Susceptible-Exposed-Infectious-Removed (SEIR) model was used to effectively predict the peaks and sizes of COVID-19 epidemiological data with sufficient fitting performance. A modified SEIR model also showed a good effect for predicting the peaks and sizes. The peak deviation of the another modified SEIR model in predicting epidemiological data in China was 3.02%. However, SEIR focuses on predicting trends in sensitive, exposed, infected, and recovered groups, rather than new daily cases. Moreover, it is necessary to comprehensively consider the changes in some parameter values as affected by the changes in epidemic policies and regional differences, such as the effective reproductive number, number of contacts, and infection rate. • A modified stacked auto-encoder was used to predict the cumulative and new confirmed cases of COVID-19 in China. The average absolute error was between 0.73% and 2.27% among the cumulative confirmed cases, and the fitting of daily new cases had room for improvement. This model involves processing a latent matrix generated by a second latent layer and forming the eigenvectors for 34 provinces and cities in China; therefore, the operation is more complicated, and the lower values in the early epidemic data will result in smaller errors. • Long-Short-Term-Memory (LSTM) is more successful than auto-encoder for sequence processing tasks such as machine translation and speech recognition. In COVID-19 trend prediction, LSTM has a deeper network structure and simple operations than the modified stacked auto-encoder. A neural network model based on LSTM was used to predict COVID-19 trend after pre-training using SARS data, and the results demonstrated that LSTM has good prospects for predicting the trend; however, the fitting effect needs to be improved. A hybrid model with an LSTM was used to predict COVID-19 trends in 11 countries, and showed accurate predictive ability. The Root-Mean-Square-Error (RMSE) from the model was between 15 to 623, and we estimated that the error was between 2% and 24%, except for the error of approximately 84% in China, but this model was sensitive to the large fluctuation in the data. In this study, to reduce the prediction errors and the sensitivity of daily data with large fluctuation for a better fitting effect, we train an improved three-layer superimposed LSTM (3*LSTM) neural network model based on single-layered LSTM, and use the weekly epidemiological data of China, five overseas countries and the world to predict the trend. We also use the model to analyse policy events and the social economic influences under the epidemic situation. Data sources and study area The study data comprised the epidemiological data of COVID-19 between January 16 and December 5, 2020 in China and the current confirmed epidemic data between January 12 and December 5, 2020 in Wuhan, Hubei Province, China, as reported by the National Health Commission of the People’s Republic of China (PRC) and the Health Commission of Hubei Province. The data of the top five countries from January 22 to December 5, 2020 worldwide were sourced from a dataset based on global COVID-19 epidemic statistics, as published by Johns Hopkins University in the United States. The five overseas countries were the United States, India, Brazil, Russia and France. The cumulative confirmed overseas epidemic data from February 15 to December 5, 2020 were retrieved from a Baidu website. The study area is shown in the Figure 1. We combined the Chinese and overseas COVID-19 data to generate the numbers of cumulative confirmed cases and cumulative deaths worldwide. In the process of training, we found that better prediction results could be obtained among data with more evident changes. Therefore, we took the Saturday data from the daily accumulated diagnostic data of China, the five countries, and the world. As shown in the Table 1, the start date of COVID-19 data in the study country or region varied with the start time of the epidemic; the deadline for the statistical data was December 5, 2020. Epidemiological data from China and the United States was used for multi-layer LSTM comparison and trend prediction. Data from Wuhan was used to compare the prediction effects of our LSTM model with that from SEIR models. Data from the United States, India, Brazil, Russia, France was used for comparing outbreaks in the five countries, and the global epidemiological data was used for forecast trends.


Introduction
The coronavirus disease 2019 (COVID-19) is brought on by infection from Severe Acute Respiratory Syndrome (SARS) Coronavirus 2, and the reports of related cases were first released by Wuhan, Hubei Province, China in December 2019 1 . The widespread COVID-19 epidemic is a serious threat, and has become one of the most challenging global catastrophes facing mankind since the Second World War 2 . On March 11, 2020, the global development of COVID-19 was assessed by the World Health Organization (WHO) as having met the characteristics of a pandemic 2 . The mortality rate of COVID-19 is estimated to be between 2% and 5%, i.e. lower than that of SARS and Middle East Respirator Syndrome 4,5 . However, COVID-19 has a higher infection rate than bat-like SARS, and its pathogenicity is between that of SARS and bat-like SARS 6 . Patients with COVID-19 generally have symptoms such as fever, dry cough, fatigue and dyspnea, the immunological function is weakened, and in severe cases, respiratory failure and death may occur [7][8][9] . The heart may be damaged and develop myocardial inflammation after recovery from COVID-19 9 . In addition, the slowdown or stagnation of national economies during the epidemic prevention period has caused a deceleration in global economic growth 11 , along with aggravating unemployment and hunger 12 . Some national leaders have been infected with the virus, thereby affecting the implementation of policies and management.
On January 15, 2020, the Chinese Center for Disease Control and Prevention (China CDC) initiated a first-level emergency response 13 . China began implementing measures since January 23, 2020 14 , and the epidemic had stabilized by March, 2020. In contrast, the current epidemic situation overseas is not very optimistic. The number of infected people continues to maintain substantial increase of tens of thousands. According to the report by the WHO on December 6, more than 200 countries, territory and regions have been invaded by COVID-19, the global epidemic has infected more than 65.8 million people, with 1523656 deaths, and the epidemic in the Europe, Americas and South-East Asia are the worst, especially in the United States and India 15 .
At present, 52 candidate vaccines have entered phase clinical trials; 13 of these have entered the third phase 16 . It is expected that vaccines may be applied by the end of 2020. In addition, countries with outbreaks have taken many measures, such as city lockdowns, banning unimportant gatherings, and business opening in advance, but the effect of the policies are unclear 17 . The president of the United States will officially transition on January 20, 2021, and new policies affecting the epidemic situation may be promulgated 18 . Therefore, the aim of this study is to predict subsequent trend according to the epidemiological data of COVID-19, and to conduct a policy analysis to summarise meaningful and positive measures. This information has great reference significance for assessing the development of the epidemic situation in China, and in five countries with serious epidemic situations 14 . According to the forecast, it will be helpful to assess the effectiveness of current policies and make flexible adjustments, for instance, reinforcement, alteration, and relaxation of restrictions.
According to incomplete statistics, there are three main methods for forecasting the epidemic development of COVID-19, as follows: • The Susceptible-Exposed-Infectious-Removed (SEIR) model was used to effectively predict the peaks and sizes of COVID-19 epidemiological data with sufficient fitting performance 14,19 . A modified SEIR model also showed a good effect for predicting the peaks and sizes 14 . The peak deviation of the another modified SEIR model in predicting epidemiological data in China was 3.02% 19 . However, SEIR focuses on predicting trends in sensitive, exposed, infected, and recovered groups 14,19 , rather than new daily cases. Moreover, it is necessary to comprehensively consider the changes in some parameter values as affected by the changes in epidemic policies and regional differences, such as the effective reproductive number, number of contacts, and infection rate. • A modified stacked auto-encoder was used to predict the cumulative and new confirmed cases of COVID-19 in China. The average absolute error was between 0.73% and 2.27% among the cumulative confirmed cases, and the fitting of daily new cases had room for improvement 20 . This model involves processing a latent matrix generated by a second latent layer and forming the eigenvectors for 34 provinces and cities in China; therefore, the operation is more complicated, and the lower values in the early epidemic data will result in smaller errors. • Long-Short-Term-Memory (LSTM) is more successful than auto-encoder for sequence processing tasks such as machine translation and speech recognition 21 . In COVID-19 trend prediction, LSTM has a deeper network structure and simple operations than the modified stacked auto-encoder. A neural network model based on LSTM was used to predict COVID-19 trend after pre-training using SARS data, and the results demonstrated that LSTM has good prospects for predicting the trend; however, the fitting effect needs to be improved 14 . A hybrid model with an LSTM was used to predict COVID-19 trends in 11 countries, and showed accurate predictive ability. The Root-Mean-Square-Error (RMSE) from the model was between 15 to 623, and we estimated that the error was between 2% and 24%, except for the error of approximately 84% in China, but this model was sensitive to the large fluctuation in the data 22 .
In this study, to reduce the prediction errors and the sensitivity of daily data with large fluctuation for a better fitting effect, we train an improved three-layer superimposed LSTM (3*LSTM) neural network model based on single-layered LSTM, and use the weekly epidemiological data of China, five overseas countries and the world to predict the trend. We also use the model to analyse policy events and the social economic influences under the epidemic situation.

Data sources and study area
The study data comprised the epidemiological data of COVID-19 between January 16 and  Figure 1.
We combined the Chinese and overseas COVID-19 data to generate the numbers of cumulative confirmed cases and cumulative deaths worldwide. In the process of training, we found that better prediction results could be obtained among data with more evident changes. Therefore, we took the Saturday data from the daily accumulated diagnostic data of China, the five countries, and the world. As shown in the

Data processing
Sequence Data Generation is a crucial step, and it is the premise of LSTM model training. The sequence data can reflect changes in single data in which no regularity can be traced. Therefore, the individual data were organised into sequential data.
As shown in the Figure 2, the individual data dimension is ( , 1), a window with a length (sequence length) of and step size of 1 was initially selected for sliding orderly through the daily data. All of the data contained in the window was considered as a data sequence, and the next day data outside the window was the corresponding label data. After the window slid through all of the daily data, the sequence data with dimension of ( − , ) was generated.

Long-Short-Term-Memory (LSTM)
LSTM is a particular type of recurrent neural network (RNN), and is mainly used to learn and overcome the long-term dependence of RNN; they both have the same basic structure 27 . The cyclic function in an RNN is implemented in the hidden layer, and the other two main components of the RNN are the input and output layers. Each cyclic unit has an input and output 21 . Except for the last layer of cyclic units, the output of each cyclic unit and the input of the next time together determine the output of the next time 28 . However, too many cyclic units may lead to the loss of previously learned rule information and the existence of long-term dependence, and can lead to the problem of gradient disappearance or gradient explosion 21 .

Figure 2. Sequence Data Generation
The weight value of the self-circulation introduced by LSTM is determined based on context, rather than fixed 21 . The internal units of LSTM include the state unit, forget gate, input gate and output gate 21,28 . The state unit is a core part, and information is transmitted from an upper-level unit through the state unit to the next cyclic unit after the internal update of the state unit, and the information is not affected by other internal calculations 28 . The sigmoid units in the gates constrain the value between 0 and 1 so as to determine which information should be saved or forgotten, thereby solving the problem of gradient disappearance, while learning long-term dependence 21,28 . In an LSTM cell, the first step is to discard unnecessary information through the forget gate. The second step is to determine the information stored in the state unit through the input Daily data (n) gate and tanh activation layer. The input gate protects the memory information in the cyclic cell from the influences of other inputs. The final output information is determined by the output gate, and the output gate protects other cells from the memory information in the cyclic cell 27,28 .
In the epidemic data, the daily new case data had a certain range of continuous jitter, which led to the change of the growth rate of cumulative cases. The fitting degree of a single-layer LSTM model is low. Moreover, the prediction effect of continuous jitter data was poor and had high sensitivity.

Improved LSTM model
To better train with continuously fluctuating data, fit unsmooth curves, extract regular information more effectively, and improve prediction capabilities, we deepened the number of LSTM layers. The established 3*LSTM model was used to train based on the COVID-19 epidemic data, and to predict the development trend.
In the 3*LSTM model, the memory carries the changing regularity transfers between the LSTM layers and LSTM internal circulation units. When the information flow enters the circulation unit ( Figure 3), the operation is as follows: In the above, the LSTM layer = 1, 2, 3, stands for the moment, , and represent the output after the activation function in the forget gate, input gate and output gate, respectively. In the cell at time of LSTM in the layer , the first step of information flow is to discard unnecessary information. The information matrix or vector comprising the output ℎ −1 of the last moment and the input of the current moment are linearly combined with the weight , and the bias offset is added. Then the linear function returns the output through the activation function . Second, the output of the activation layer in the input gate is obtained by information matrix [ℎ −1 , ], multiplied by the weight , and added with the bios offset . The activation function is used to determine which information needs to be saved or updated, and then a vector containing the updated information ̃ is generated through the ℎ activation layer. The updated cell state is obtained by adding * −1 which represents unnecessary information to be discarded, and * ̃ which represents the information to be updated. Finally, whether the information will be output is selected by the activation layer . The information value tanh( ) (in the range -1 to 1 after ℎ layer) is obtained and multiplied by the output of active layer to return the final output ℎ for this LSTM cell.
As shown in Figure 4, 0 , 1 , …, represents sequence data generated from individual data. The three different coloured layers represent the expanded cyclic units of the three LSTM layers. Each LSTM cell has an input and output ℎ.
The final output of the upper LSTM layer is used as the input to the next LSTM layer. The output of the first LSTM layer, and second LSTM layer is [ℎ 0 , ℎ 1 , … , ℎ ], ( = 1, 2), which represents the set of outputs from each cyclic unit in that layer.
The third LSTM layer outputs the final ℎ 3 , i.e. the output value of the last cyclic unit of the layer. Then it determines the dimension to be output through the dense layer. Finally, the linear activation layer is used to fit the output data and label data to further reduce the error. The output data of the model continues to be trained to predict data trends. indicates the sequence data, ℎ represents the output of the LSTM recurrent unit. The blue is the first LSTM layer, orange is the second LSTM layer, and yellow is the third LSTM layer.

Assessment metrics
To evaluate the effect of our model, the RMSE was used to represent the fitting degree between the output (prediction data) and label data of each sequence data after 3*LSTM model training. The RMSE function is as follows: In the LSTM Model, denotes the number of sequential data, and ℎ and represent the output of each sequence after LSTM training and the corresponding label data, respectively. The smaller the value of RMSE, the better the prediction effect, which means that the error between prediction data and label data is smaller.

Training environment
The deep learning environment of our experiment was mainly built based on the Ubuntu system environment, which combines Keras and Tensorflow for training with the support of a GPU. The detailed parameters are showed in Table 2.

Comparison of LSTM models with different layers
We used the COVID-19 weekly confirmed case data from China and the United States to evaluate the LSTM model. To compare the prediction effects from the different amount of LSTM layers, we trained these networks with the time steps of 3, 5, 7, 9 and 11 weeks. The error rate of these models exceeds '1' when the time step is 3 or 5 weeks. Therefore, we do not show these results.

Comparison of 3*LSTM model and Susceptible-Exposed-Infectious-Removed (SEIR) model
We used current confirmed cases in Wuhan from February 12, 2020 to December 5, 2020, and the forecast date ended on January 16, 2021. The current confirmed cases in Wuhan refer to confirmed cases excluding deaths and rehabilitated people, and we regard it as the same parameter as the 'infected people' parameter in the SEIR model. We took 3*LSTM model for training with a time step of 33 days and 3000 epochs, and the RMSE was 30.49. According to the SEIR model, the population in Wuhan, Hubei Province, China is divided into four groups based on disease stage: susceptible (S), exposed (E), infected (I), and recovery (R) 19

Comparison of the epidemic situations between China and the US
The first COVID-19 case of in China began to show symptoms on December 1, 2019 30 . After the outbreak, the Chinese government issued a timely emergency response: laboratories expanded their capacities for nucleic acid detection, the 'Huoshenshan', 'Leishenshan', and Fangcang shelter hospitals were established for accommodating more patients, medical observation was conducted on close contacts, the source of infection was investigated, people remained at home as much as possible, and wearing masks, avoiding gatherings, and keeping a physical distance of 1 meter when going out were strictly required 17,31 . The China CDC issued prevention guidelines for people of different ages and places to strengthen the awareness of safety 32 . Under the guidance of national policies and the common resistance of people across the country, the epidemic has been promptly controlled, and the number of confirmed cases has gradually decreased. The epidemic situation has entered the stage of continuous prevention and control, and the resumption of work and school has been flexibly promoted 33 .
On January 20, 2020, the United States recorded its first COVID-19 patient 34 . The initial measure after the outbreak in the United States was to temporarily ban entry of non-Americans who had visited China in the past 14 days 35 . However, the number of infected people increased, owing to flows of people at home and abroad 36 . The CDC proposed that residents wear masks in public places to reduce transmission on April 3 37 , but many people ignored this advice, and previously unknown challenges existed, owing to a campaign of mass protests and campaign rallies without protective measures 3636 . In addition, the 'opening up' measures promoted by the American government led to increases in the quantity of hospitalizations 38 Table 4 and   Table 5.
We used cumulative data from China and the United States for training. Owing to the lack of weekly data, and to better modeling effects, we used odd numbers from 3 to 35 weeks as the time steps for independent training. The optimal prediction effect was obtained when the time step of cumulative cases was 33 weeks. As shown in Figure 6a, the epidemic in China has subsided since March, the amount of daily new confirmed cases has gradually been maintained below 50. The COVID-19 deaths in China have been basically flat since late April and have remained at 4634 since May 6, and the weekly mortality rate has dropped to 0 (Error! Reference source not found.b, Figure 7). In Figure 7b, 1290 death cases were increased after data revision in mid-April, 2020, resulting in abnormal peak value. The quantity of cumulative confirmed cases in the United States has been rising sharply after surpassing China on March 26, 2020, the growth rate has not shown a decline, and deaths in the United States are increasing by thousands a week ( Figure 8). The weekly mortality rate in the United States has surpassed that of China since late April, 2020; fortunately, the mortality has shown a downward trend ( Figure 9). In particular, mortality in the United States has gradually declined after reaching its maximum in May; thia may be related to the policy of wearing masks. The data in Table 6 shows that the difference between the predicted and the actual values is less than 0.4%, indicating that the model has good predictive fitting performance. (2) The promotion of resumption of work and school 33 .

Top five countries in the global cumulative diagnosis
The weekly cumulative confirmed case data from five countries are trained through the 3*LSTM model. These five countries are the US, India, Brazil, Russia, and France. From the perspective of epidemic prevention and control, in some countries, measures were not strictly enforced, or the severity of the epidemic was ignored so as to restart the economy (e.g. the early release of blockade restrictions in some states of the United State) 52 . Brazil implemented a social distancing policy, but some people neglected to wear masks at rallies 53 , and professional football matches were held in Rio, Brazil on June 18, 2020 54 .
The Indian government implemented the 'city closure' measures at the end of March, but some areas have been re-blocked, because of an outbreak rebound caused by the increase of outdoor activities after a gradual unsealing 55       The time steps used for cumulative confirmed cases and deaths were 33. Figure 8 shows the forecasts for the five countries.
Except for France, the cumulative confirmed cases and deaths of the other four countries have no evident flattening trend.
Russia has more cumulative confirmed cases and fewer deaths than France, but the numbers of cumulative deaths is expected to exceed France in early January, 2021. In Table 7, the error rate between predicted and actual values is less than 1.4%.
According to the forecast, the cumulative cases of the worst outbreak (in the United States) will exceed 17.8 million by the end of 2020 and 19.9 million in mid-January, 2021. The epidemic situations in the US, India and Brazil remain high-risk. Figure 9 shows the cumulative and weekly mortality rates, based on actual and predicted cumulative confirmed cases and deaths in the five countries. The cumulative mortality rate in India, Russia and France is expected to maintain an increase within 0.1%, whereas those in the United States and Brazil are estimated to decline slowly. The continuously declining weekly mortality rate in the United States may rise after 3 weeks. The weekly mortality rate in the other four countries is forecast to rise slightly. It is expected that the higher weekly mortality rate in France next week may be caused by lower number of new confirmed cases. In addition, the weekly mortality rate in France may reach 5.6% on the third Saturday of January, 2021.

Trend prediction of global diagnosis
The 3*LSTM model was used to predict epidemic trends by training the cumulative confirmed cases for the global week from February 15 to December 5, 2020. Table 8 shows the results from model fitting and prediction, and Figure 10 displays the prediction generation diagram. According to the forecast, the quantity of cumulative diagnoses worldwide continues to rise, but the speed will gradually slow down. By January 16, 2021, the number of global cumulative confirmed cases may surpass 84.58 million, and deaths may reach 1.85 million. In Figure 11, we calculate the mortality of COVID-19, based on actual and predicted cumulative confirmed cases and deaths. The mortality of global COVID-19 patients has been declining after reaching a peak (7.12%) in May, and this trend is expected to continue. Nevertheless, the weekly mortality rate may rise. LSTM, Long-Short-Term-Memory.

Discussion
According to the experimental results, it can be considered that the LSTM model can understand the regular pattern of epidemic data, has good learning capabilities, and fits training data more accurately. The prediction accuracy of 3*LSTM model is better than that of the other LSTM models in this study. Our model shows that the error rate of China, five countries and the world is less than 1.4%. The fitting degree of the 3*LSTM model is better than the that of SEIR model among the COVID-19 data from Wuhan, and the 3*LSTM model can combine multiple real-time data for prediction. In addition, the model applies the rules learned from the training data in the trend prediction stage, especially the cumulative case data of the top five countries and the world, and the predicted results conform to the general trends shown by the data.
The high numbers of the cases, frequent changes in the slope of the curve, and the scarcity of training data will affect the fit of the model. This is an abnormal phenomenon in that when cumulative confirmed cases are predicted in a longer time step, the predicted value is consistent with the true value, but the predicted value decreases during the trend forecast phase.
One possibility is that the model overly emphasises the decreasing trend of slope value when modelling the gradually gentle data curve, causing the prediction curve to decline. Although our model performs well for data fitting, it is only a forecast for the current policy situation, and cannot predict sudden increases in data at a later stage owing to policy easing or other activities. Thus, the prediction results of the model are for reference only.
Based on summarizing positive and negative measures during the epidemic (Table 9), we conclude that the timely release of prevention and control policies by the government is the most important. The movement and exposure of people within a country should be reduced by sealing off cities, closing entertainment venues and the prohibiting activities. The establishment of border inspection and cruise ship bans is beneficial to preventing the entry of people from other countries with the virus.
The establishment of the temporary hospitals can improve the ability to accommodate quarantined patients. The release of prevention and control guidelines contributes to guiding people to take correct self-protection methods. Improving the production of medical supplies and the ability to test for COVID-19 ensures that patients can be detected and treated promptly, and increases recovery rates. Policy relaxation and economic restarts should be conducted after the epidemic situation is relatively stable, rather than prematurely. Large-scale activities with crowd gatherings should be cancelled as much as possible, owing to the existence of a large potential infection rate. At the resident level, some countries have promulgated corresponding measures, but citizens have not strictly implemented them. Some citizens resist the 'closing the city' measure, for they think it conflicts with personal freedom. Some people have negative emotions about the severe epidemic and determination of them to protest together is insufficient, affecting the policy response. It is essential to actively respond to the relevant prevention and control policies issued by the state, and not to violate the prohibition without authorisation.
Although the current epidemic situation in some countries is gradually stable, keeping vigilant, paying attention to personal protection, and reducing long-distance travel remain important.
In the view of the actual situation and prediction, Chinese epidemic is in a stable state, whereas the United States is firmly in the first place worldwide; moreover, the dynamic trends of aggravation of the COVID- 19  severe economic damage, market economic imbalances, factory or company closures, significantly increased unemployment, displacement of people, and increased hunger. To avoid gatherings of people, the elections in the United States will also be affected. In summary, the most effective measure is to self-protection, and the most basic protection is to wear a mask and keep a physical distance. In addition to effective prevention and control policies, only active self-protection can minimise the chance of infection. To prevent the global crisis caused by the continued deterioration of the epidemic, economic restarts, and resumptions of work and school should be avoided as much as possible before the epidemic has stabilised.

Conclusions
In this work, an LSTM-based neural network model was proposed to predict the dynamics of COVID-19 in China, five countries with severely epidemics, and the world. The error rate of our model was less than 1.4%. We demonstrated that the 3*LSTM model, with good fitting performance, is superior to other methods, and that the trend predictions basically conform to the actual current situations. According to forecasts, if the present measures are maintained, the amount of cumulative confirmed cases in the globe will exceed 84.58 million by January 16, 2021, with a global declining mortality rate.
Under the analysis of measures in China and the five other countries, untimely and inappropriate government policies, lax regulations, and insufficient public cooperation are the reasons for the aggravation of the epidemic situations. A summary is provided, as follows. First, it is vital for the government to take appropriate and proactive measures in time. Simultaneously, the active cooperation of the public, and faith in fighting the epidemic together is needed to promote the implementation of the measures for obtaining good effects. The first and most essential approach is to wear masks and keep a physical distance.
Other active policies should be advocated, such as the release of emergency responses, lock-downs of cities and closures of entertainment venues, prohibitions of gatherings as much as possible, the establishment of temporary hospitals, border inspections, bans on cruise ships, prevention and control guidelines, the vigorously production of medical materials, and improvements in the ability to detect COVID-19. Second, the epidemic situations of the United States, India, and Brazil are not encouraging. Opening in advance, neglecting protective measures, having gatherings, and insufficient medical resources have provided great potential for infection. Third, in countries where the epidemic situation has not stabilised or is growing significantly, according to the stage of the epidemic, it is necessary to adjust policies flexibly, increase prevention and control efforts, expand the intensity of nucleic acid testing, deploy medical supplies, and avoid large-scale gatherings. In countries where the epidemic has gradually stabilised, to prevent the recurrence of the epidemic after the resumption of production, personal protection should be maintained, together with disinfection at the places of work, study and entertainment in time, along with body temperature monitoring of entry and exit personnel.
Only with the use of a vaccine will we be able to curb the spread and multiplication of the virus. However, the timing of emergence, effectiveness, and side effects of the vaccine are unknown. Thus, people must wear a mask and maintain a physical distance to take precautions against COVID-19 for the next two years, even with widespread vaccination.