Machine learning for electric energy consumption forecasting: Application to the Paraguayan system.

This article focuses on short-term electric energy consumption forecasting using a time series approach applied to data from a Paraguayan electricity distribution company collected over three years. The study compares various machine learning (ML) and statistical methods—including linear regression (LR), random forests (RF), generalized boosted regression models (GBM), multilayer perceptron (MLP), and long short-term memory networks (LSTM)—to predict seven-day ahead electricity consumption using different historical window sizes. Meteorological data (temperature, humidity, wind speed, atmospheric pressure) were incorporated to assess their impact on prediction accuracy. Results indicate that the deep learning LSTM model generally outperforms other methods, especially when meteorological data are included, allowing for accurate predictions with smaller historical windows; LR performed better than RF and GBM but was outperformed by LSTM and MLP. The study highlights the benefit of integrating weather data in forecasting models and suggests future exploration of advanced deep learning architectures for improved energy demand prediction.