COVID-19 Forecasting Based on Local Mean Decomposition and Temporal Convolutional Network

Since the outbreak of coronavirus disease 2019 (COVID-19) has resulted in a dramatic loss of human life and economic disruption worldwide from early 2020, numerous studies focusing on COVID-19 forecasting were presented to yield accurate predicting results. However, most existing methods could not provide satisfying forecasting performance due to tons of assumptions, poor capability to learn appropriate parameters, etc. Therefore, in this paper, we combine a traditional time series decomposition: local mean decomposition (LMD) with temporal convolutional network (TCN) as a general framework to overcome these shortcomings. Based on the particular architecture, it can solve weekly new confirmed cases forecasting problem perfectly. Extensive experiments show that the proposed model significantly outperforms lots of state-of-the-art forecasting methods, and achieves desirable performance in terms of root mean squared log error (RMSLE), mean absolute percentage error (MAPE), Pearson correlation (PCORR), and coefficient of determination (R-2). To be specific, it could reach 0.9739, 0.8908, and 0.7461 on R-2 when horizon is 1, 2, and 3 respectively, which proves the effectiveness and robustness of our LMD-TCN model.