Previsão a curto prazo de preço de energia utilizando modelo híbrido com sarimax e xgboost

Abstract

Efficient short-term forecasting of electricity purchase prices is a crucial challenge, particularly in the Brazilian context, characterized by high price volatility and a dynamic energy market. The peculiarities of the Brazilian electricity sector, such as dependence on renewable sources, seasonal climatic variations, and the growing complexity of integrating distributed systems, make this task even more demanding. Most available methods focus on direct time series modeling without fully leveraging the structural decomposition of data to enhance predictability. This study proposes a hybrid model (SARIMAX-XGBoost) for short-term forecasting, combining the seasonal autoregressive integrated moving average model (SARIMAX) with the extreme gradient boosting algorithm (XGBoost). SARIMAX captures structural components such as trends, seasonality, and external variable effects, while XGBoost models residuals and unexplained complex patterns. The final forecast is obtained by integrating the outputs of both models. Validation was performed using real market data from Brazil, covering historical series of electricity prices. Experimental results demonstrate that the SARIMAX-XGBoost model effectively captures market volatility, providing accurate and stable predictions, outperforming advanced methods such as artificial neural networks and purely XGBoost-based approaches in both accuracy and stability.