Hybrid Prediction Model Based on Decomposed and Synthesized COVID-19 Cumulative Confirmed Data

Abstract

Since 2020, COVID-19 has repeatedly arisen around the world, which has had a significant impact on the global economy and culture. The prediction of the COVID-19 epidemic will help to deal with the current epidemic and similar risks that may arise in the future. So, this paper proposes a hybrid prediction model based on particle swarm optimization variational mode decomposition (PSO-VMD), Long Short-Term Memory Network (LSTM) and AdaBoost algorithm. To address the issue of determining the optimal number of modes K and the penalty factor (α) in the variational mode decomposition (VMD), an adaptive value for particle swarm optimization (PSO) is proposed. Specifically, the weighted average sample entropy of the relevant coefficients is utilized to determine the adaptive value. First, the epidemic data are decomposed into multiple modal components, known as intrinsic mode functions (IMFs), using PSO-VMD. These components, along with policy-based factors, are integrated to form a multivariate forecast dataset. Next, each IMF is predicted using AdaBoost-LSTM. Finally, the prediction results of all the IMF components are reconstructed to obtain the final prediction result. Our proposed method is validated by the cumulative confirmed data of Hubei and Hebei provinces. Specifically, in the case of cumulative confirmation data, the coefficient of determination (R2) of the mixed model is increased compared to the control model, and the average mean absolute error (MAE) and root mean square error (RMSE) decreased. The experimental results demonstrate that the VMD–AdaBoost–LSTM model achieves the highest prediction accuracy, thereby offering a new approach to COVID-19 epidemic prediction.