Accepted_test

COVID-19 caused by a novel coronavirus has continued to pose as a serious public health risk in 2024. The mutation of the virus has reached seasonal frequency, which requires additional resources of the medical system. Enough data on the dynamics of COVID-19 in the world since 2020 allows the use of statistical analysis and machine learning methods to build forecasts of epidemic time series. Classical machine learning and time-series approaches do not allow for probabilistic forecasts or have strong assumptions on the form of the future distribution of the target variable. We used an unsupervised deep learning algorithm consists in two neural networks: the generator G generates new samples z of data close to the true data x and the discriminator D distinguishes generated samples of data from the true samples x with condition y.

We apply Condition Generative Adversarial Networks (CGANs) for probabilistic forecasting of the spread of COVID-19 in Saint-Petersburg using 15 processed epidemic time series, such as new tested, diagnosed, hospitalized, critical (requiring a ventilator), immunity cases, deaths, self-isolation index in considered region as well as new diagnosed COVID-19 cases in the World and its combinations. We construct 1, 3 and 5-days prediction of newly COVID-19 diagnosed and compare results with classical SIR-type epidemic model and full-connected neural network. The percentage of successful hits in the confidence interval constructed by CGAN for 3-sigma rule for 1 day prediction is 75% and for 5 days prediction is 67%.