A Comparative Analysis of Neural Network Architectures for Energy Theft Detection

Electricity theft is a pressing, prevalent, and pervasive issue for both utility companies and their consumers, especially in densely populated developing countries like Pakistan. Conventional theft detection methods require manual inspection or intervention, and are not just expensive but also inef

Project Title

A Comparative Analysis of Neural Network Architectures for Energy Theft Detection

Project Area of Specialization

Artificial Intelligence

Project Summary

Electricity theft is a pressing, prevalent, and pervasive issue for both utility companies and their consumers, especially in densely populated developing countries like Pakistan. Conventional theft detection methods require manual inspection or intervention, and are not just expensive but also ineffective in identifying and isolating electricity thieves. The result is utilities incurring financial losses and consumers (even those who do not commit electricity theft) being deprived of electric power in areas where non-technical losses are exceedingly high.

The project described in this proposal aims to solve these problems through a data-driven approach to electricity theft detection. Concretely, the project’s goal is to use Deep Learning to model electricity consumption patterns for theft detection using a variety of neural network architectures. This project will use advanced metering infrastructure (AMI) power consumption data to train different kinds of neural networks to identify electricity theft. A comparative analysis of the aforementioned neural networks will then be used to identify and deploy the neural network with the highest accuracy and lowest false positive rate on a cloud platform to act as a backend for a theft detection web application for utility companies.

Project Objectives

Accurate Electricity Theft Detection for Pakistan: We aim to identify and build a neural network model/architecture that can accurately and efficiently identify electricity theft using advanced metering infrastructure (AMI) data. Concretely, this means the neural network model should have both high sensitivity and high specificity. This will empower utility companies to pinpoint instances of electricity theft using AMI data with greater accuracy at a fraction of the cost of existing methods.

Comparative Neural Network Analysis: Our primary goal is to investigate, compare, and contrast the performance of various neural network architectures for the purpose of electricity theft detection using AMI data. Through a systematic testing process, our project will identify the neural network architecture that is best suited for electricity theft detection on AMI data.

Deploying a Neural Network Backend: Once the best neural network architecture has been identified and its resulting model adequately trained, the latter is to be deployed on a cloud platform. The model can thus be used as a backend solution for utility companies to identify electricity theft using their own AMI data.

Mastering Deep Learning:The objective of our FYP is to develop a thorough understanding of Deep Learning theory, tools, and techniques in the context of an end-to-end, real world power engineering problem. As electrical engineering undergraduates, we have had limited exposure to Deep Learning in an academic environment. Consequently, this project is an excellent opportunity for us to develop our expertise in AI and Deep Learning, empowering us to become better problem solvers in a world on the brink of the 4th Industrial Revolution.

Project Implementation Method

Acquiring Data

The first step in the project’s implementation is to procure AMI energy consumption data that will be used to train and test our neural networks. The ideal data set

• will be sourced from a Pakistani Distribution Corporation (DISCO) or utility so the NN will generalize well to Pakistani electricity consumption patterns.

• will have sufficiently large number of training examples, with each being a time series consisting of sufficiently many features such as hourly kWh consumption measurements.

• will be labelled: any time series data corresponding to electricity theft will be labelled as such.

If such a data set cannot be procured, our project will alternatively

• create a simulated data set that mimics Pakistani electricity consumption patterns for both paying consumers as well as electricity thieves.

• Use an AMI data set provided by the State Grid Corporation of China as done by Zheng et al (2017).

Literature Review

Simultaneously, an extensive literature review will be carried out to determine the types of Neural Network architectures to be tested, the correlation between specific electricity consumption patterns and electricity theft, and the data preprocessing required to maximise neural network prediction accuracy.

Neural Network Design and Development

The data set will be divided into a training set and a test set, with the former being used to train a variety of neural network architectures (CNN, RNN, etc.) using the Python-based TensorFlow library. The test set will be used to evaluate the performance of each neural network architecture using a suite of performance metrics and methods such cross validation, AUC, specificity/sensitivity. Data will be cleaned, processed, and transformed as required based on the neural network model being investigated.  All models will be tested on either Google Cloud Platform or AWS.

Final Model Selection and Deployment

The model with that gives the best performance across the aforementioned suite of metrics will then be optimised and deployed to a cloud platform. Utility companies will then be able to use the model as a black box to perform electricity theft detection using their own AMI data. The minimum viable interface will be an API that can be accessed through a command line tool.

Benefits of the Project

Smarter Energy Theft Detection: Our neural network will provide a non-invasive, and autonomous solution for electricity theft detection in smart grids and power distribution networks. Prevalent theft detection methods in Pakistan involve manual inspection and intervention, and are therefore neither efficient nor cost-effective. The neural network will allow DISCOs and utility companies to improve the process of identifying and taking action against electricity thieves, making it smarter - cheaper, faster, and more convenient.

For Utilities: It is estimated that up to 40% of distribution expenditure of DISCOs and utilities in developing countries is due to non-technical losses or electricity theft. For a country like Pakistan with an endemic energy crisis and increase in demand outpacing generation capacity growth, minimising non-technical losses in a cost-effective manner is an utmost priority. Utilities can combine their AMI data with our cloud-based theft detection model to pinpoint instances of electricity theft amongst their consumers at a fraction of the cost of conventional methods. This will translate into lower operating costs for utilities, better energy efficiency, and improved cost per unit generation, streamlining the process of energy generation and distribution.

For Consumers: Operating costs associated with targeted manual anti-electricity theft operations are very high. As such, utility companies will rarely penalise individual consumers suspected of electricity theft and instead increase load shedding/power penalties for all consumers in areas affected by electricity theft. We feel this is an ineffective incentive structure: it does very little to permanently deter electricity theft and does not offer any benefits to consumers who do not commit electricity theft. Our neural network approach makes it possible to identify targeted, specific electricity theft detection, which will empower utilities to take action against only those consumers that actually contribute to NTL. This could eliminate unnecessary load shedding and similar penalties imposed on areas with electricity theft.

Incentive for AMI Proliferation: Our NN will offer a cost-effective, software-based solution to targeted NTL reduction, with all the benefits this will entail. Consequently, it will act as an added incentive for DISCOs and utilities to invest in and popularize AMI, bringing Pakistan closer to achieving its goal of turning its metropolises into smart cities.

Technical Details of Final Deliverable

Our project’s final deliverable is a neural network based on the Python Deep Learning framework Tensorflow that will be deployed on either Google Cloud Platform (GCP) or Amazon Web Services (AWS). The neural network will be able to process energy consumption time series data for different consumers obtained through AMI. The NN will then identify consumers suspected of electricity theft along with a corresponding confidence level. The NN will also be accompanied by a report that details our investigation and findings which informed our decision to use a specific neural network architecture and any required data preprocessing.

Final Deliverable of the Project

Software System

Type of Industry

Energy

Technologies

Artificial Intelligence(AI), Big Data

Sustainable Development Goals

Sustainable Cities and Communities

Required Resources

If you need this project, please contact me on contact@adikhanofficial.com