Scholarlink

<p>The COVID-19 pandemic, caused by the SARS-CoV-2 virus, has had a major impact on global health and economies. It is crucial to accurately forecast the number of confirmed COVID-19 cases and deaths to allocate medical resources effectively and formulate government policies. This study aims to improve COVID-19 time series forecasting by developing the universal COVID-19 time series patterns (UCovid-19P) model, which uses deep learning techniques and leverages transfer learning to optimize the accuracy of predictions. The proposed method uses a two-stage approach: first, pretraining on a comprehensive COVID-19 dataset to acquire universal patterns, and then transfer learning on specific country data. Various deep learning architectures, such as long short-term memory (LSTM), gated recurrent units (GRUs), bidirectional LSTM, and bidirectional GRUs, are implemented. Evaluation metrics like root mean square error (RMSE), mean absolute error, mean absolute percentage error (MAPE), symmetric mean absolute percentage, root mean squared log error, and explained variance (EV) are used to evaluate the model’s performance in forecasting confirmed cases and deaths, both in one-step and multi-step scenarios, across 16 countries spanning three continents. Experimental findings demonstrate that the UCovid-19P model significantly outperforms baseline models, achieving an average MAPE improvement exceeding 70%. Additionally, both the UCovid-19P and baseline models demonstrate robustness by effectively explaining variance, as evidenced by their EV scores. Comparative analysis against state-of-the-art models reveals that the UCovid-19P model achieves minimal errors, as evaluated by MAPE, compared to LSTM and GRU architectures, and similar performance to models employing attention mechanisms and transformers. The UCovid-19P model provides a robust and scalable framework for COVID-19 forecasting. By integrating transfer learning, the proposed model enhances its capability to predict future time series data. This enhancement is crucial for preparing and strategizing responses to future pandemics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Universal COVID-19 time series patterns model for confirmed and death cases forecasting using transfer learning

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Paisit Khanarsa,
  • Satanat Kitsiranuwat

摘要

The COVID-19 pandemic, caused by the SARS-CoV-2 virus, has had a major impact on global health and economies. It is crucial to accurately forecast the number of confirmed COVID-19 cases and deaths to allocate medical resources effectively and formulate government policies. This study aims to improve COVID-19 time series forecasting by developing the universal COVID-19 time series patterns (UCovid-19P) model, which uses deep learning techniques and leverages transfer learning to optimize the accuracy of predictions. The proposed method uses a two-stage approach: first, pretraining on a comprehensive COVID-19 dataset to acquire universal patterns, and then transfer learning on specific country data. Various deep learning architectures, such as long short-term memory (LSTM), gated recurrent units (GRUs), bidirectional LSTM, and bidirectional GRUs, are implemented. Evaluation metrics like root mean square error (RMSE), mean absolute error, mean absolute percentage error (MAPE), symmetric mean absolute percentage, root mean squared log error, and explained variance (EV) are used to evaluate the model’s performance in forecasting confirmed cases and deaths, both in one-step and multi-step scenarios, across 16 countries spanning three continents. Experimental findings demonstrate that the UCovid-19P model significantly outperforms baseline models, achieving an average MAPE improvement exceeding 70%. Additionally, both the UCovid-19P and baseline models demonstrate robustness by effectively explaining variance, as evidenced by their EV scores. Comparative analysis against state-of-the-art models reveals that the UCovid-19P model achieves minimal errors, as evaluated by MAPE, compared to LSTM and GRU architectures, and similar performance to models employing attention mechanisms and transformers. The UCovid-19P model provides a robust and scalable framework for COVID-19 forecasting. By integrating transfer learning, the proposed model enhances its capability to predict future time series data. This enhancement is crucial for preparing and strategizing responses to future pandemics.