As the COVID-19 pandemic continues to ravage the world, it is critical to assess the COVID-19 risk timely on multi-scale. To implement it and evaluate the public health policies, we develop a machine learning assisted framework to predict epidemic dynamics from the reported infection data. It contains a county-level spatio-temporal epidemiological model, which combines spatial cellular automata (CA) with time sensitive-undiagnosed-infected-removed (SUIR) model, and is compatible with the existing risk prediction models. The CA-SUIR model shows the multi-scale risk to the public and reveals the transmission modes of coronavirus in different scenarios. Through transfer learning, this new toolbox is used to predict the prevalence of multi-scale COVID-19 in all 412 counties in Germany. A t-day-ahead risk forecast as well as assessment of the non-pharmaceutical intervention policies is presented. We analyzed the situation at Christmas of 2020, and found that the most serious death toll could be 34.5. However, effective policy could control it below 21thousand, which provides a quantitative basis for evaluating the public policies implemented by the government. Such intervening evaluation process would help to improve public health policies and restart the economy appropriately in pandemics.

随着 COVID-19 大流行继续肆虐世界，在多尺度上及时评估 COVID-19 风险至关重要。为了实施它并评估公共卫生政策，我们开发了一个机器学习辅助框架，以根据报告的感染数据预测流行病动态。它包含一个县级时空流行病学模型，该模型将空间细胞自动机（CA）与时间敏感未诊断感染清除（SUIR）模型相结合，并与现有的风险预测模型兼容。CA-SUIR模型向公众展示了多尺度风险，揭示了冠状病毒在不同场景下的传播方式。通过迁移学习，这个新工具箱用于预测德国所有 412 个县的多尺度 COVID-19 的流行情况。介绍了 t 日前风险预测以及非药物干预政策的评估。我们分析了2020年圣诞节的情况，发现最严重的死亡人数可能是34.5。但是，有效的政策可以将其控制在21000以下，这为评估政府实施的公共政策提供了定量依据。这种干预性评估过程将有助于改善公共卫生政策并在流行病中适当地重启经济。