COVID-19 belongs to a vast group of viruses rapidly disseminating worldwide. The substantial surge in infected individuals has resulted in a strain on medical facilities across the globe. Consequently, precise predictions of case numbers and fatalities can assist governments and various organizations in preplanning response strategies, alleviating the burden on the healthcare system. Deep learning introduces a fresh approach to addressing emerging coronavirus outbreaks. In this study, we present an innovative hybrid framework that combines singular spectrum analysis (SSA), empirical fourier decomposition (EFD), and temporal convolutional network (TCN) for the prediction of COVID-19. To leverage COVID-19 time series data to its fullest extent, we have devised a residual frequency domain segmentation technique by merging SSA and EFD. Through this approach, the initial time series, characterized by intricate fluctuations, can be restructured into subsequences featuring more conspicuous patterns of change. TCN-based models utilize dilated causal convolution and residual connections to encompass both immediate and prolonged dependencies within time series data. To showcase the predictive capabilities of our proposed model, we conduct short, medium, and long-term forecasts for weekly new confirmed cases and deaths, respectively. Experimental reveal that our model outperforms the other seven baseline models in terms of root mean squared logarithmic error (RMSLE), mean absolute percentage error (MAPE), coefficient of determination (R2), and pearson correlation coefficient (PCC).