A Comparative Study on Deep Learning Models for COVID-19 Forecast

Abstract

The COVID-19 pandemic has led to a global health crisis with significant morbidity, mortality, and socioeconomic disruptions. Understanding and predicting the dynamics of COVID-19 are crucial for public health interventions, resource allocation, and policy decisions. By developing accurate models, informed public health strategies can be devised, resource allocation can be optimized, and virus transmission can be reduced. Various mathematical and computational models have been developed to estimate transmission dynamics and forecast the pandemic’s trajectories. However, the evolving nature of COVID-19 demands innovative approaches to enhance prediction accuracy. The machine learning technique, particularly the deep neural networks (DNNs), offers promising solutions by leveraging diverse data sources to improve prevalence predictions. In this study, three typical DNNs, including the Long Short-Term Memory (LSTM) network, Physics-informed Neural Network (PINN), and Deep Operator Network (DeepONet), are employed to model and forecast COVID-19 spread. The training and testing data used in this work are the global COVID-19 cases in the year of 2021 from the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University. A seven-day moving average as well as the normalization techniques are employed to stabilize the training of deep learning models. We systematically investigate the effect of the number of training data on the predicted accuracy as well as the capability of long-term forecast in each model. Based on the relative L2 errors between the predictions from deep learning models and the reference solutions, the DeepONet, which is capable of learning hidden physics given the training data, outperforms the other two approaches in all test cases, making it a reliable tool for accurate forecasting the dynamics of COVID-19.