Outlier Detection and Correction in Smart Grid Energy Demand Data Using Sparse Autoencoders

The implementation of smart grids introduces complexities where data quality issues, particularly outliers, pose significant challenges to accurate data analysis. This work develops an integrated methodology for the detection and correction of outliers in energy demand data, based on Artificial Neural Network autoencoders. The proposed approach is submitted across multiple scenarios using real-world data from a substation, where the influence of the variation in the number of outliers present in the database is evaluated, as well as the variation in their amplitudes on the functioning of the algorithms. The results provide an overview of the operation as well as demonstrate the effectiveness of the proposed methodology that manages to improve some indices achieved by previous works, reaching accuracy and F-score superior to 99% and 97%, respectively, for the detection algorithm, as well as a square root mean squared error (RMSE) and a mean absolute percentage error (MAPE) of less than 0.2 MW and 2%, respectively.