Detecção de outliers em séries temporais de carga com redes neurais recorrentes

An important aspect of the load curve provided by the National Electric System Operator is the ability to analyze variations in energy consumption over time. Hourly demand, influenced by seasonality, serves as an indicator of expected energy consumption and the proper functioning of the National Interconnected System. While visually identifying outliers in the time series of hourly load is possible, this approach becomes laborious and imprecise. Therefore, it is preferable to automatically detect outliers, considering seasonal patterns. In this study, a model of Artificial Neural Network with Long Short-Term Memory cells was employed to identify outliers based on seasonality. Historical energy consumption data collected over time in the Southeast/Central-West subsystem were used. The Interquartile Range (IQR) technique was employed for data labeling. The ANN was trained based on labeled data for the years 2020 and 2021, and subsequently tested with 2022 data. The results were positive, achieving precision and recall metrics of 98% and 96%, respectively.