Forecasting SARS-CoV-2 in the peruvian regions: a deep learning approach using temporal convolutional neural networks

• Autores: Aguilar I., Luis, Ibáñez-Reluz, Miguel, Z. Aguilar, Juan C., Zavaleta-Aguilar, Elí W., Aguilar, L. Antonio
• Formato: artículo
• Fecha de Publicación: 2021
• Institución: Universidad Nacional de Trujillo
• Repositorio: Revistas - Universidad Nacional de Trujillo
• Lenguaje: inglés
• OAI Identifier: oai:ojs.revistas.unitru.edu.pe:article/3698
• Enlace del recurso: https://revistas.unitru.edu.pe/index.php/SSMM/article/view/3698
• Nivel de acceso: acceso abierto
• Materia: Deep Learning; Forecasting; SARS-CoV-2; Temporal Convolutional Neural Networks; Time Series Data; Aprendizaje profundo; Previsión, SARS-CoV-2; Redes Neuronales Convolucionales Temporales; Datos de Series de Tiempo

The SARS-CoV-2 pandemic had taken the world by surprise since its discovery on December 2019, causing major losses worldwide. In this work, a deep learning model was developed to predict and forecast the daily SARS-CoV-2 cases on the Peruvian regions. The data used belongs to the open covid–19 data set, sourced by the Health Ministry of Peru (MINSA). The data set includes the periods from March 03, 2020 to March 16, 2021. A holdout approach was used, creating a training and validation data splits. Using the validation set, a temporal convolution neural network (TCN) composed by five layers was developed. The model was design to predict a mean tendency alongside with a prediction interval. To find the best hyper parameter configuration, a Bayesian approach was applied over the validation set. The TCN model was trained using the optimal configuration. Once trained, the model was able to predict the different SARS-CoV-2 trends present in the regions. Next, a forecast was performed beyond the available data, using a window of 15 days ahead (March 17 to March 31, 2021) for each region. Forecast results suggested a continued trend for all the regions, except Lima. The model performance was evaluated using the MAE, MAD, MSLE and RMSLE metrics on the test period, showing training to validation metrics improvements of 14.534, 3.123, 0.042, 0.047 respectively.