Artificial Neural Network Techniques for Energy Theft Detection in AMI

Authors: Diwakar Agrawal, Asst. Prof. Mamta Devi Sharma, Associate Prof. Ravi Kumar Hada

Certificate: View Certificate

Abstract

The success of using renewable energy sources is driving the biggest change to the power grid in a long time. This change is happening from centralised control of the electricity supply to an infrastructure that is smart and not centralised. On the other hand, when different parts of a power grid become more connected to each other, these parts become more vulnerable to cyber attacks, fraud, and software problems. In recent years, a lot of progress has been made in cyber-physical security, like the ability to detect physical tampering, as well as more traditional information security solutions, like encryption. Traditional information security solutions, on the other hand, can\'t handle all of the problems that cyber threats pose, because digital electricity meters can have software bugs and hardware problems. As a result of making electricity meters digital, many security problems that had been solved in the past, such as electricity theft, have come back as IT problems. For these new problems, we need new ways to find out about them that are based on data analysis, machine learning, and making predictions. Rapid changes in statistical methods, which are similar to techniques used in machine learning, have led to a rise in interest in ideas that can model, predict, or extract load information, such as that provided by a smart meter, to spot early signs of tampering. Anomaly Detection Systems can find tampering techniques by looking at statistical deviations from a normal behaviour that has already been set. This method is widely accepted as a good way to find abuse patterns that were not known before. This study gives a number of anomaly detection algorithms that use power measurements to make it easier to find tampered electricity meters early on. Time series prediction and probabilistic models have been used to create and test algorithms with detection rates of more than 90%. Several things were taken into account in this process. One of the things that was added was the study of complex threads, such as behaviors that are similar to each other. Other contributions include an analysis of the different kinds of data that are already available, the creation of metrics and methods of aggregation that make it easier to find specific patterns, and so on. This work helps us learn more about the important features and typical behaviour of electric load data, as well as find evidence of tampering and, more specifically, energy theft.

Introduction

The internet of things (IoT) and artificial intelligence (AI) are two cornerstone technologies enabling smart cities, and have been interacting with each other into an organic ecosystem. In the smart grid, smart meters and various sensors are widely used to increase the two-way communication capability. Combined with the advanced metering infrastructure (AMI), they enable energy companies to obtain real-time voltage, current, active power, reactive power, energy usage and other measurements from the smart meters deployed at user homes [1],[2]. Recently, smart meters are shown to be vulnerable to cyber physical attacks in the smart grid due to their insecure and distributed network and physical environment [3]–[5]. One serious threat is energy theft attacks, which cost more than 25 billion dollars every year to the energy companies [6]. Such an attack aims to pay less by attacking user meters to tamper with the energy usage sent to energy company. Another severe threat is privacy violation. As smart meters collect real-time energy usage that may reveal user’s habits and behavior at home, the user privacy concern will be raised if the collected data is not well protected [7]. For example, if the user’s daily energy consumption is low, it may imply that the user is not at home [8]. Thus, such privacysensitive information must be protected from unauthorized access. To disclose the usage for theft detection and to hide the usage for privacy preservation are conflicting goals. We aim to address both theft detection and privacy preservation in this work. A number of works have been conducted for energy theft detection in the smart grid. Some used the classificationbased support vector machine (SVM) technique to classify the normal and attack samples from the energy usage database [9]–[11]. In addition, matrix decomposition [12], linear regression [13] and state estimation [14] can be used to analyze the data for energy theft detection. However, these approaches cannot be applied to cases with massive amounts of data. [15] proposed a wide and deep convolutional neural network model to analyze energy theft behavior of individual users. In our paper, we additionally study the energy thef

Conclusion

To detect any kind of theft whether conventional or data attack, distributed totalization metering is employed in conjunction with artificial intelligence. In distributed totalization metering is employed in conjunction with Artificial Intelligence. In distributed totalization metering, a energy company owned main meter is placed at distribution end (feeder/distribution transformer) before preventing electricity at consumers premises. In theory, the sum of all the meters at consumer premises fed by that particular distribution and should be equal to the reading of main meter. However practically, there may be some difference due to T&D (Transmission & Distribution) losses, load fluctuations, power factor issues, calibration issues etc. Here artificial intelligence comes into play, which monitors the reading of the main meter as well as the consumer meters, & gets trained during normal operation, whenever energy theft occurs, & accumulates, artificial intelligence detects the anomaly & raises on alert for energy theft , so that consumer premises, feed by that distribution end can be inspected.