Anomalies In Commercial Power Consumption

Abstract

To promote sustainable power management techniques, anomaly detection in power consumption data is essential. This research provides a thorough methodology for identifying anomalies in smart meter data by combining exploratory data analysis and innovative machine learning approaches. The suggested approach obtains high precision in identifying deviated consumption patterns by utilizing convolutional neural networks (CNNs) in conjunction with autoencoders and Mahalanobis distance computations. Experimentation showed that temperature, day of the week, and time of day all had a significant impact on complicated trends that highlighted the dynamics within power use. Granular anomaly localization, which accurately identified anomalous time points within aberrant days, was made possible by the incorporation of reconstruction errors from autoencoders and Mahalanobis distances. With the test dataset, the system effectively identified 622 anomalies, proving its efficacy. The study identifies several constraints and difficulties, including problems with data quality, the requirement for integrating domain expertise, and the significance of optimizing dynamic thresholding methods, even though the outcomes are encouraging. Subsequent avenues for research include hybrid techniques that integrate several machine learning models, validate the framework at the appliance level for targeted power management, and incorporate domain knowledge. All things considered, by accurately identifying anomalies, this study helps to optimize power management systems and opens the door for more costeffective, sustainable, and power-efficient projects