A Data Mining-Based Dynamical Anomaly Detection Method for Integrating with an Advance Metering System

TL;DR

This paper presents a real-time dynamical anomaly detection system for smart meters, combining supervised and unsupervised machine learning techniques with adaptive thresholds to improve early detection of power consumption anomalies.

Contribution

It introduces a novel dynamic anomaly detection system using both Light Gradient Boosting and autoencoders with adaptive thresholds for real-time power monitoring.

Findings

Effective detection of anomalies in real-world power data

Adaptive thresholding improves detection accuracy over static methods

System enables early warning for power consumption irregularities

Abstract

Building operations consume 30% of total power consumption and contribute 26% of global power-related emissions. Therefore, monitoring, and early detection of anomalies at the meter level are essential for residential and commercial buildings. This work investigates both supervised and unsupervised approaches and introduces a dynamic anomaly detection system. The system introduces a supervised Light Gradient Boosting machine and an unsupervised autoencoder with a dynamic threshold. This system is designed to provide real-time detection of anomalies at the meter level. The proposed dynamical system comes with a dynamic threshold based on the Mahalanobis distance and moving averages. This approach allows the system to adapt to changes in the data distribution over time. The effectiveness of the proposed system is evaluated using real-life power consumption data collected from smart metering systems. This empirical testing ensures that the system's performance is validated under real-world conditions. By detecting unusual data movements and providing early warnings, the proposed system contributes significantly to visual analytics and decision science. Early detection of anomalies enables timely troubleshooting, preventing financial losses and potential disasters such as fire incidents.