Short-term forecasts and long-term mitigation evaluations for the COVID-19 epidemic in Hubei Province, China

Qihui Yang; Chunlin Yi; Aram Vajdi; Lee W Cohnstaedt; Hongyu Wu; Xiaolong Guo; Caterina M Scoglio

The predictions of COVID-19 cases in Wuhan city from February 22 nd to March 1 st were conducted. Every predictive simulation was run 250 iterations. The removing rate we used was 1/7 and the infection rate was estimated by EnKF. The model noise and measurement noise was tuned by state of art. Fig. 1 shows the distributions of the predicted numbers of COVID-19 cases based on data 1, 2 and 3 days before. The results of the EnKF prediction give two rangesthe weak one from minimum to maximum (the solid black line) which can include all the reported real data (the green circle); the stronger one from the lower quartile to the upper quartile (the blue box) which also covers most of the real numbers. The red line inside the box indicates the median of the samples. We can see that most medians are very close to the corresponding real numbers. Taking the median as our predicted number gives the relative error shown in Fig. 1(d) . The mean absolute percentage error (MAPE) for 1-day-ahead prediction is 0.19%; the MAPE for 2-day-ahead prediction is 0.32%; the one for 3-day-ahead is 0.62%. The performance of 1-day-ahead prediction is the best and the one of 3-day-ahead is the worst. When n-day-ahead prediction is conducted, the results from (n-1)-day-ahead is used, so the errors will accumulate.