detection of unauthorized electricity consumption …...university unit name here detection of...

17
University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department of Electrical and Computer Engineering Mississippi State University

Upload: others

Post on 04-Jun-2020

4 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Detection of Unauthorized Electricity Consumption using Machine Learning

Bo Tang, Ph.D.Department of Electrical and Computer 

EngineeringMississippi State University

Page 2: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Outline• Advanced Metering Infrastructure (AMI) in Smart Grid

• Unauthorized Electricity Consumption (UEC)

• Detection of UEC using Machine Learning Algorithms

• Preliminary Simulation Results

• Future Work and Challenges

Page 3: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Overview of Advanced Metering Infrastructure

• AMI:• Comprised of state‐of‐the‐art 

electronic/digital hardware and software.

• Enable detailed, time‐based data measurements and their transmissions.

• Benefits:• System operation benefits• Customer service benefits• Financial benefits 

Source: Electric Power Research Institute

Page 4: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Unauthorized Electricity Consumption• Ways of unauthorized electricity consumption (energy theft)

• Taking connections directly from distribution line• Grounding the neutral wire• Inserting some disc to stop rotating of the coil• Hitting the meter to damage the rotating coil• Interchanging input output connections

• Difficult to check these issues with AMI.• It is estimated that utility companies lose more than $25 billion every 

year due to energy theft around the world.

Page 5: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Techniques for UEC Detection• Statistical approaches

• Hypothesis test• Learning‐based Approaches

• SVM, Neural Networks, Fuzzy classification, ARMA‐GLRT • State‐based approaches

• Sensor monitoring, RFID monitoring, Mutual inspection, State estimation‐based.

• Game theory‐based approaches

Page 6: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Detection of UEC with Machine Learning

Fig. 1 Basic procedure for learning‐based energy‐theft detection

Page 7: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Anomaly Detection• Anomaly is a pattern that does not conform to the expected behavior.• Also referred to outliers, exceptions, surprises, novelty, etc. 

• General Steps for Anomaly Detection:• Build a profile (or pattern) of the normal behavior• Use the normal profile to detect anomalies

• Anomalies are observations whose characteristics  differ significantly from normal profile.

Page 8: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Proposed Method• Clustering the normal data to build multi‐modal profiles• K‐mean clustering algorithms• Silhouette (cohesion and separation) measure is 

to used to determine the number of patterns (clusters)

• Apply distance‐based (neighbors‐based) anomaly detection approaches.

Page 9: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Relative Density-based Outlier Score (RDOS)

• Local Kernel Density Estimation

Page 10: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Relative Density-based Outlier Score (RDOS)

• Relative Density‐based Outlier Scores

• Anomalies detections  p 

Page 11: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

RDOS: Theoretical Properties

Page 12: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Experimental Results• Datasets:

• Smart energy data from the Irish Smart Energy Trial, including hourly electricity usage reports of Irish homes in 2009 and 2010. 

• Synthetic unauthorized energy consumption• Seven types of energy theft are generated. 

Page 13: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Experimental Results

Page 14: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Experimental Results• Seven distance‐based anomaly detection algorithms:

• Relative Density‐based Outlier Score (RDOS) • Local Outlier Factor (LOF)• Local Density Factor (LDF)• Flexible Kernel Density Estimates (KDEOS)• Influenced Outlierness (INFLO)• Mutual k‐nearest neighbor (MNN) • Indegree Number(ODIN)

Page 15: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Experimental Results• Top 3 anomaly detection algorithms: AUC (area under ROC curve)

Theft‐Type 1st  2nd  3rd

1 RDOS(0.88) LDF(0.84) INFLO(0.80)

2 LDF(0.79) RDOS(0.76) ODIN(0.72)

3 RDOS(0.93) MNN(0.86) ODIN(0.81)

4 RDOS(0.94) ODIN(0.86) INFLO(0.85)

5 RDOS(0.96) LDF(0.90) INFLO(0.88)

6 LDF(0.95) RDOS(0.93) ODIN(0.88)

7 RDOS(0.94) LDF(0.89) KDEOS(0.8)

OA RDOS(0.92) LDF(0.85) INFLO(0.80)Example of ROC curve

Page 16: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Future Work• Unauthorized energy consumption detection

• Advanced detection methods• Real‐life data sets

• Privacy preservation in AMI• Cloud computing and big data

Page 17: Detection of Unauthorized Electricity Consumption …...University Unit Name Here Detection of Unauthorized Electricity Consumption using Machine Learning Bo Tang, Ph.D. Department

University Unit Name Here

Questions?