Enhancing COVID-19 forecasts with a lightweight multi-head depthwise separable convolution network

To address the challenge of forecasting COVID-19 with limited data, this paper presents CDSCnet (Chunking Depthwise Separable Convolution network). It is a lightweight model designed for small datasets, employing three fixed convolution heads to capture long-term dependencies without relying on recurrent units. Through a comprehensive comparative analysis with existing models, including CNN-LSTM variants, using COVID-19 datasets from 7 countries, such as India, Brazil, and the United States, we demonstrate the superior performance of CDSCnet. Across both the 8 smooth datasets and the 3 high-noise datasets, CDSCnet achieved optimal prediction accuracy, attaining a maximum MAE reduction exceeding 50% against the reproduced results in Spain task. These experimental results confirm that CDSCnet effectively captures the dynamics of epidemic spread, proving its utility as a reliable decision-support tool for COVID-19.