Análise do uso de transformers na predição de séries temporais Campina Grande – PB.

Abstract

The Transformer-based architecture, developed to tackle natural language processing machine learning problems, has expanded into other areas such as time series forecasting. Various models, based on regression or decision trees, are employed in this field, such as ARIMA, XGBoost, and Prophet, for instance. Each approach has its own specificities in terms of accuracy and computational efficiency, necessitating studies to determine the best one to adopt in different scenarios. In this context, this work aims to present comparative and analytical results of different approaches for time series forecasting. To do so, a comparative study was conducted analyzing the usage of the Transformer-based architecture and its performance against traditional models, using data related to energy consumption records from the Federal University of Campina Grande (UFCG). Through this analysis, it was possible to assess the approaches in terms of accuracy and computational efficiency, determining whether the complexity of a more sophisticated approach like Transformers for time series forecasting proves superior to other popular approaches utilized for this purpose.