DRFD: Deep Learning-Based Real-time and Fast Detection of False Readings in AMI

Abstract

Smart meters, installed at customers’ apartments, frequently send their power consumption readings to the system operator in the advanced metering infrastructure (AMI) network. These readings are used for energy management, load estimation, and billing. Nonetheless, malicious customers, who aim to lower their bills illegally, launch electricity theft cyberattacks by breaching their meters and reporting lower readings. These reported false readings are toxic to the grid’s reliability and performance because they are used in energy management, and hence causing financial losses and inefficient use of resources. Existing solutions present in the literature aim at securing billing only because they are designed to detect false readings in real-time. Therefore, the SO may continue to make use of these false readings for energy management and load monitoring for a long time until the detection is done. In this paper, we propose real-time detection of false readings using deep learning. We first create malicious and benign datasets generated from a sliding window and use them to train different deep learning models. The best-performing model is then trained on various ratios of the false readings. In comparison with the existing daily and weekly electricity theft detection methodologies that require 144 and 1,008 readings, respectively, our detector can identify false readings after transmitting a few false readings (about 20).