COVID-19 Forecasting in India through Deep Learning Models

Abstract

COVID-19 pandemic has affected global economy severely with major impact towards medical and agricultural systems. Due to infections spread complexity there has been unreliability in renowned computational and statistical models. There has also been uncertainty in modeling due to inherent lack of data collection and reporting activities. As a result of this, situation needs to be revisited with latest data sources and comprehensive forecasting models. In order to achieve this, deep learning models like long short-term memory (LSTM) networks and its variants have been instrumental towards modeling temporal sequences. In this paper, different versions of LSTM networks like bidirectional LSTM, encoder-decoder LSTM, hierarchical bidirectional LSTM and hierarchical encoder-decoder LSTM have been used towards forecasting COVID-19 spread in different Indian states. The states are selected considering COVID-19 hotpot regions with respect to infection rates. A comparative study is performed with states having contained or peak infection stages. It has been observed that with LSTM networks and its variants convincing long-term forecast results are achieved which can be utilized in other countries.