A Deep Learning Approach to Forecast SARS-CoV-2 on the Peruvian Coast

ABSTRACT

The current spreading of the SARS-CoV-2 pandemic had put all the scientific community in alert. Even in the presence of different vaccines, the active virus still represents a global challenge. Due to its rapid spreading and uncertain nature, having the ability to forecast its dynamics becomes a necessary tool in the development of fast and efficient health policies. This study implements a temporal convolutional neural network (TCN), trained with the open covid-19 data set sourced by the Health Ministry of Peru (MINSA) on the Peruvian coast. In order to obtain a robust model, the data was divided into validation and training sets, without overlapping. Using the validation set the model architecture and hyper-parameters were found with Bayesian optimization. Using the optimal configuration the TCN was trained with a test and forecasting window of 15 days ahead. Predictions on available data were made from March 06, 2020 until April 13, 2021, whereas forecasting from April 14 to April 29, 2021. In order to account for uncertainty, the TCN estimated the 5%, 50% and 95% prediction quantiles. Evaluation was made using the MAE, MAD, MSLE, RMSLE and PICP metrics. Results suggested some variations in the data distribution. Test results shown an improvement of 24.241, 0.704 and 0.422 for the MAD, MSLE and RMSLE metrics respectively. Finally, the prediction interval analysis shown an average of 97.886% and 97.778% obtained by the model in the train and test partitions. © 2021, Springer Nature Switzerland AG.