Multivariate time series prediction of Covid-19 daily new cases in Indonesia based on Deep Learning: Unlocking the role of effective reproduction number (Rt)

Abstract

Abstract To date, COVID-19 and its variants have been among the greatest hindrances for humanity. This disease is spreading rapidly and almost all parts of the world are currently exposed to it. The ability to understand and simultaneously predict the dynamics of daily confirmed cases of this disease is essential to prevent and mitigate the impact of the pandemic. This study investigates the use of Deep Learning (DL), including Deep Feedforward Neural Networks (DFNN), Long Short-Term Memory (LSTM), a one-dimensional convolutional neural network (CONV1D), and Gated Recurrent Units (GRU), to predict daily confirmed cases of Covid-19 in Indonesia by taking into account as many as 25 variables (predictors) as inputs. Variable filtering was also performed to identify the predictors with the best weight. Extreme Gradient Boosting (XGBoost) regression is used for this purpose. Some statistical analyses were also carried out to increase our understanding of the data before modelling. The performance of the algorithm was assessed using several metrics, including Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Scaled Error (MASE). MASE is a measure of MAE relative to the baseline model. The results showed that DL using two predictors, the number of daily confirmed cases and the Rt (effective reproduction number) value, had the highest performance and was able to predict the number of daily confirmed cases 13 days ahead. Adding more variables deteriorates DL performance.