Predição de séries temporais da covid-19: uma avaliação do uso dos modelos suavização exponencial, ARIMA, MLP & LSTM

Abstract

In this master’s degree dissertation, it will be discussed how the predictive models ARIMA, LSTM, MLP and Exponential Smoothin were developed and implemented to predict time series of confirmed cases and deaths from COVID-19, to assess which among these obtains the best result. COVID-19 is a disease caused by the coronavirus called SARS-CoV-2, which has resulted in a large number of infected people globally. According to the WHO, more than 305 million people are estimated to be infected worldwide. As it was necessary to use reliable data to carry out the predictions, the database used for the development of this dissertation is in the public domain and was provided by the Johns Hopkins University. Time series data of confirmed cases and deaths from Brazil, India, Italy and the United States of America were compared and selected to make predictions. About the predict models, the Long Short-Term Memory neural network is capable of learning long sequences of observations to make predictions. Besides this, the Multi-Layer Perceptron is a neural network with one or more hidden layers with an undetermined number of neurons. In addition, the ARIMA is an autoregressive integrated moving average. Finally, Exponential Smoothing is a highly accurate prediction model for smoothing time series data. Therefore, after carrying out the training and testing of each of the models, the performance evaluation was carried out with the root-mean-square error (RMSE) method and based on the results of the implemented models for the prediction of data referring to the confirmed cases and deaths from the COVID-19 pandemic, it was possible to evaluate that the ARIMA model had the best performance among the others.