Aplicação de redes neurais recorrentes para previsão de séries temporais

Abstract

Time series forecasting is a topic that has become increasingly popular in recent decades, thanks to its diverse applications in various areas of knowledge and the constant growth in the amount of data that is produced every second by an infinity of devices around the world. There is a wide range of forecasting methods, each with specific properties that must be considered when choosing a suitable method for each application scenario. In recent years, machine learning models have gained prominence and established themselves as serious competitors to classical statistical models, especially recurrent neural networks, which have a kind of memory capable of retaining past information and using it to predict future values. This work proposes to investigate the performance of recurrent neural models against more traditional approaches in time series modeling in two different application scenarios, based on forecast accuracy metrics. The first scenario investigated the effectiveness of these models in approximating the results obtained in computing heart rate variability indices through data coming from photoplethysmography sensors to those achieved by electrocardiogram sensors. In a second application scenario, the performance of these models was evaluated in forecasting the volume of calls received by the call center of an electricity distribution company. In both scenarios, the performance of recurrent neural models was superior to statistical models, indicating that they can be an appropriate choice for time series forecasting in several areas.