A deep learning approach for prediction of SARS-CoV-2 cases using the
weather factors in India
Abstract
Advanced and accurate forecasting of COVID-19 cases play a crucial role
in management of hospital facility, policy decision, logistic support,
and economy of the country. Artificial Intelligence (AI) techniques have
proved its capability in time series forecasting of the non-linear
problems. The present study assessed the relationship between weather
parameters and COVID-19 cases and found the specific humidity have
strong positive association, maximum temperature have negative and
minimum temperature have positive association in most of the states in
India. Further, we have developed a weather integrated LSTM (long short
term memory) models for advanced (1-14 days) forecasting of the COVID-19
cases over different states in India. To achieve the goal we have
utilized the humidity and temperature time series data along with the
COVID-19 confirmed cases data (1st April-30th June 2020) to optimise the
LSTM model in univariate and multivariate modes. The optimised models
are utilized to forecast the COVID-19 cases for the period 1st July,
2020 to 31st July 2020 with 1 to 14days lead time. The results shows
that the univariate LSTM model (past COVID-19 input) have reasonably
good skill (Relative Error < 20%) in short range forecast
(1day lead) for most of the selected states, whereas the skill is
degraded with the medium and long range forecast. The major finding of
the current study is that the medium range (1-7days) forecasting skill
is enhanced in some of the states with the weather integrated
multivariate LSTM models. The states (Maharashtra, Gujarat, Rajasthan,
Madhya Pradesh, Haryana, and Punjab) located in West and North West
India region, humidity play a key role in enhancement of medium range
forecasting skill of the LSTM model. It is also observed that the states
located in high humid regions (Kerala, Tamil Nadu, and West Bengal)
temperature plays a key role in model enhancement.