Self-Driven Time Series Forecasting of COVID-19 Pandemic Using Reservoir Computing

The time series forecasting of COVID-19 epidemic which has brought a serious health and economic crisis worldwide has become very imperative. Our study aims to provide a relatively simple deep learning way to model, forecast, and evaluate the time evolutions of COVID-19 pandemic. In this paper, we built a data self-driven reservoir computing (RC) model to predict the evolution of COVID-19 time series. Specifically, the self-driven reservoir was created to form the closed feedback loop through replacing the input with the output. According to the different temporal evolution trend (gentle or dramatical) of COVID-19 time series, we proposed a data self-adaptive prediction scheme that the self-driven reservoir could autonomously adapt its network parameters to data changes in order to improve prediction accuracy. The prediction results which involved COVID-19 temporal evolutions of the multiple countries around the world indicated the excellent prediction performances of the proposed model in comparison with some of the main AI prediction models from literatures (e.g. RNN, LSTM, GRUs, VAE) at the same time scale. Moreover, we found that the model parameters of the self-driven reservoir had a great impact on the prediction performance, as a result, we could select appropriate parameters to improve the prediction performance. Finally, because time dependent length of COVID-19 time series forecasting is varying around inflection points, we analyzed the influence of short-term memory capacity of the self-driven reservoir on prediction performance around inflection points. The analysis showed that the reservoir could adaptively adjust short-term memory capacity according to change trend of data to improve prediction performance. Our prediction results can be used as proposals to help governments and medical institutions to formulate pertinent precautionary measurements to prevent further spreads.