Multistep Forecasting of New COVID-19 Cases Based on LSTMs Using Bayesian Optimization

Abstract

The multistep prediction of new Corona Virus Disease (COVID-19) cases plays a vital role during the epidemic control period, and the Long Short-Term Memory (LSTM) based time series analysis model is the most frequently used among many prediction methods. But whether it is the cumulative error of the multistep prediction or the instability of the new case data of the COVID-19 make the performance of LSTM in this task not so good. In this paper, we selected three countries with more severe COVID-19 epidemics—India, Russia, and Chile, to predict new cases in the next 15 days with different multistep LSTM network models, and use Bayesian Optimization to explore the optimal hyperparameter space. The results show that: a) the performance of Recursive Prediction LSTM is the best (Mean Absolute Percentage Error, MAPE was reduced to 14.88%, 6.46%, and 16.31% for the three countries respectively), Encoder Decoder LSTM is second (15.52%, 19.61%, 19.87%), and the effect of vector output LSTM is the worst (23.55%, 26.82%, 19.57%); b) there are obvious extremely poor areas in the hyperparameter space, and the Bayesian Optimizer can focus on the good areas to avoid cost of tuning parameters based on bad hyperparameters; c) the data of new cases of COVID-19 in different countries have great differences in the hyperparameter expectations for the model. The bad area of hyperparameters and different expectations are likely to be one of the reasons why the COVID-19 data of different countries is hard to train jointly.