A Review of anomaly detection techniques in advanced metering infrastructure

Bulletin of Electrical Engineering and Informatics
Vol. 10, No. 1, February 2021, pp. 266~273
ISSN: 2302-9285, DOI: 10.11591/eei.v10i1.2026  266
Journal homepage: http://beei.org
A review of anomaly detection techniques in advanced metering
infrastructure
Abbas M. Al-Ghaili1
, Zul-Azri Ibrahim2
, Syazwani Arissa Shah Hairi3
, Fiza Abdul Rahim4
,
Hasventhran Baskaran5
, Noor Afiza Mohd Ariffin6
, and Hairoladenan Kasim7
1,2,4
Institute of Informatics and Computing in Energy, Universiti Tenaga Nasional, Malaysia
3,5,7
College of Computing & Informatics, Universiti Tenaga Nasional, Malaysia
6
Department of Computer Science, Faculty of Computer Science & Information Technology,
University Putra Malaysia, Malaysia
Article Info ABSTRACT
Article history:
Received Dec 4, 2019
Revised Feb 28, 2020
Accepted May 4, 2020
Advanced metering infrastructure (AMI) is a component of electrical
networks that combines the energy and telecommunication infrastructure to
collect, measure and analyze consumer energy consumptions. One of
the main elements of AMI is a smart meter that used to manage electricity
generation and distribution to end-user. The rapid implementation of AMI
raises the need to deliver better maintenance performance and monitoring
more efficiently while keeping consumers informed on their consumption
habits. The convergence from analog to digital has made AMI tend to
inherit the current vulnerabilities of digital devices that prone to cyber-attack,
where attackers can manipulate the consumer energy consumption for their
benefit. A huge amount of data generated in AMI allows attackers to
manipulate the consumer energy consumption to their benefit once they
manage to hack into the AMI environment. Anomalies detection is a technique
can be used to identify any rare event such as data manipulation that happens
in AMI based on the data collected from the smart meter. The purpose of this
study is to review existing studies on anomalies techniques used to detect
data manipulation in AMI and smart grid systems. Furthermore, several
measurement methods and approaches used by existing studies will be addressed.
Keywords:
AMI
Anomaly
Data manipulation
Smart meter
This is an open access article under the CC BY-SA license.
Corresponding Author:
Abbas M. Al-Ghaili,
Institute of Informatics and Computing in Energy,
Universiti Tenaga Nasional,
43000 Kajang, Selangor, Malaysia.
Email: abbas@uniten.edu.my
1. INTRODUCTION
In the energy sector, the utilization of electric meters was initially applied to industrial
and commercial users because of the requirement to have more advanced data rates and progressively
granular charging data demands [1]. The usage bit by bit extended to all end-user classes to accommodate
a large number of customers. Automated meter reading (AMR) has been used to collect meter data by
utilizing one-way communication. In recent years, advanced metering infrastructure (AMI) has been
intensively built along with the evolution from the conventional electrical grid to the increasing growth
of the smart grid. AMI will collect, calculate and analyze the consumer power consumption data and then
transmit this information from a smart meter to a data collector center which then forwarded to a head-end
system on the utility part. Figure 1 shows the common components of AMI, the transmission of data will start

Bulletin of Electr Eng & Inf ISSN: 2302-9285 
A review of anomaly detection techniques in advanced metering infrastructure (Abbas M. Al-Ghaili)
267
from the end-user home appliance smart meter to the meter data management system (MDMS) where data
analytics and measurement of the end-user consumption for billing and management purposed.
Figure 1. Advance metering infrastructure component
The rapid implementation of AMI rises the need to deliver better maintenance and monitoring more
efficiently while keeping consumers informed on their consumption habits. The convergence from analog
to the digital meter has made AMI tend to inherit the current vulnerabilities of digital devices that prone
to cyber-attack, where attacker may gain benefit by manipulating consumers’ energy consumption [2].
Telvent, a smart grid software company in 2012 announced that their network was hacked by a group of
hackers that stole their data and infected their network with malware. Even this attack is considered a small
scale attack but a larger scale attack can give a catastrophic impact that could trigger an electrical blackout
and a total disruption to utility communication systems. In the current situation, according to [3], data
produced by smart meter in AMI MDMS when tampered could lead to anomaly pattern which demands an
anomaly detection techniques (ADTs) to detect them. As data produced in AMI is quite huge, finding
the technique required to locate anomalies in huge amounts of data is just one of the several challenges that
must be overcome in protecting the smart grid environment. In this study, a number of criteria have been
applied to identify the relevant existing studies. One of these criteria is that the study must be between 2001
and 2019. Papers related to AMI and anomalies detection methods are the main and major focus in this
review. The most recent published papers from conference proceedings and journals articles are cited in order
to highlight only the latest applied cyber-attacks on energy sector-related smart meter data that deal with
energy consumption for users. Further logical operators have been also applied to keywords during
the searching process in several digital libraries. This paper is structured as follows: Section 2 will discuss
the anomaly detection parameters. The discussion of selected studies is provided in section 3. Section 4
presents the conclusion of this work.
2. ANOMALY DETECTION PARAMETER
The past literature review has analyzed how the ADT is used in AMI and smart grid environment.
There are several attacks were introduced to attack the dataset in the past literature paper. The ADT is used
to detect the attack in the dataset. In order to evaluate the ADT, accuracy, computation time and trust
parameters were developed.
2.1. Accuracy
Anomalies detection involves the process of detecting abnormal activities from a large volume of
data, which in the AMI situation is the data supplied from the end-user smart meter to the utility provider
power consumption database. In his report [4] stated that in Great Britain, the smart meter rolled out a project
that began in 2011 and will be expected to end in 2020 to install 50 million new smart meters to 30 million
homes and non-domestics sites. With the high volume of data collection in the smart meter raised
the concerned on accuracy in detecting malicious intention to inject false data in the smart meter
consumption accuracy is used to measure the performance of the different machine learning algorithms to be
as accurate as possible.
As mention by [5] accuracy is an acceptable classified instance used in a model, while precision
is the true positive (TP) instances between all instances that considered by a model as positive. The results of
both false positive (FP) and false negative (FN) should be checked for fraud detection. In his work [5] has
reported that utility companies in the United States (US) lost around 6 billion dollars per year because of
energy theft activities. This leads to the demand for effective and accurate anomalies detection to prevent
such attacks as the smart meter still can be considered as new technology. The pattern and data behavior still
unpredictable which can cause high false-positive in anomalies detection. However, there are samples of

 ISSN: 2302-9285
Bulletin of Electr Eng & Inf, Vol. 10, No. 1, February 2021 : 266 – 273
268
methods applied to concern the dynamical behavior of energy reads and measures retrieved from smart meter
data e.g., [6] or to concern the load monitoring to reduce energy consumption utilizing smart meters [7].
Many techniques have been exploited to monitor the behavior of data or the load or specifically to detect
anomalies e.g., Neural network or machine learning [8]. It is critical to review the previous works with
the accuracy parameter used to measure the performance of different machine learning algorithms with ADT.
An extensive research studies retrieved from literature will be reviewed and identified to measure
the proportion of accurately grouped anomalies to all instances classified as anomalies.
2.2. Computation time
Time is a highly accurate parameter that can be used to measure the delay between the processing of
actual data and tampered/intruded data. Time is instrumental to show an explicit view of anomalies when
the fabricated data are compared with the actual data in a graphical view. The delay difference between
the actual data and fabricated or attacked data can be a good indicator to identify anomalies in a system.
Time-series forecasting is one of the efficient ways of detecting anomalies based on time. Any data which
is associated with time is termed as time series. Time series data can be on a daily, hourly, monthly or yearly
basis too [9]. Time series forecasting can solve typical time series forecasting problems such as demand
estimation and sales forecasting and also prepare for future needs by estimating anomalies with the current
data. Time plays a major part in anomaly detection for utility providers. Monitoring electrical consumption
over fine-grained time intervals makes utility providers detect anomalies easily. The classification algorithm
discussed in [5] detects different types of attack by having a time window for observation and detection.
In another example, authors in [10] are using time periods or time intervals in their classification based
technique to identify energy theft. This technique for energy-theft is defined as load profile classification
where the power usage of customers is recorded over a period of time. By using time as a variable,
any abnormal energy usage pattern can be distinguished from normal energy usage patterns. The use cases
mentioned above show the reliability and usefulness of time in detecting anomalies.
2.3. Trust
Trusted computing group (TCG) has defined trusted computing as to improve the overall security,
privacy, and trustworthiness of a variety of computing devices in a computing environment [10].
Trust is a metric involving nodes in a system, that referred to as the certainty of a node within the capacity,
dependability, and consistency of other nodes. There are 2 types of trust which are direct trust and indirect
trust [11]. Direct trust is dependent on viewing a node directly. Whereas indirect trust is based on reputation
which monitoring of one node by another node on success rate for duration of time. The threat level
is inversely proportional to the trust level. Lower trust level mean that threat level is high. This is why trust
is a reliable parameter in anomaly detection. This concept is applied in [12] to justify the threat impact and to
establish trust confidence on devices located in a Smart Grid.
Authors in [13] have used trust as a metric to evaluate the nodes’ behaviors and reliability in
a network. The nodes’ trustworthiness is measured with values ranging from 1 to 0. Malicious behaviors are
indicated by lower values. However, the handling of trust as an important parameter in certain systems has to
be improved. Authors in [14] have mentioned that user trust in the deep neural network-based intrusion
detection system (DNN-IDS) is imperative and high accuracy only is not enough. DNN-IDS’s transparency
hindering nature has to be reduced to be more communicative. This will eventually increase the level of user
trust. Trust also plays an important role in anomaly detection for cloud providers. Cloud service providers
(CSP) can select suitable cloud services observe the condition of CSP when performing service level
agreement (SLA) by anomaly detection algorithm. At that point refreshes CSP trust values in real-time and
dynamically alters the volume of request sent to the CSP according to their suitablility [15]. Trust is a very good
indicator for detecting anomalies, and the above-mentioned use cases are a good example of demonstrating this.
3. RESULTS AND DISCUSSION
In [5], a new feature-engineering framework using the finite mixture model (FMM) clustering
algorithm and genetic programming (GP) is proposed. The study created a practical model-agnostic
feature-engineering framework for fraud-related NTL detection in AMIs. The irish CBT [16] dataset is used
from more than 4000 households over 18 months and collected the energy consumption every 30 minutes.
There are 6 type of attacks presented in this study: (1) all the sample multiple by the random chosen
coefficient, (2) on-off attack where the consumption is set to zero during interval, (3) it multiple
the consumptions by the irregular factor that changes after some time, (4) reports a variable in time
arbitrary division of every day means utilization, (5) day by day mean utilization is always reported,
and (6) the aggregate total of power utilization is accurately presented. In this work, the GP will process

269
the information sources of the time-series meter readings of clients and train the dataset and their soft group
affiliations. Extra variable that can assist in clarifying any anomalies in user demand patterns will be
produced that combine with the time-series meter reading users in the training dataset. The performance
result for the accuracy is between 0.684 until 0.811 while the computation time is between 10.224 min until
82.113 min in detecting new types attacks such as zero day and theft events. Authors in [17] have utilized
a random tuning of hyper-parameters in Tensorflow using programming language Python to detect electricity
theft and have proposed a recurrent neural network (RNN)-based power theft detection to prevent cyber-
attacks. This model utilized the time arrangement based on clients' power consumption to create a gated
recurrent unit (GRU) which can learn customer power consumption patterns. Hence, this study was done
using irish smart energy [18] real dataset of 5000 users from 2009 to 2010 which accumulated more than 536
days. The dataset contained customers meter readings for every 30 minutes. Each customer will produce in
total 25,728 reports and the daily report for 200 customers with 107,200 days is recorded before they launch
an attack. The result shows that the RNN-based detector has a better detection rate and a lower false
acceptance rate. As such this technique only focus on the time-based pattern only.
Huang [19] suggested artificial neural networks (ANNs) detect energy theft. ANNs are
a collection of neurons that associated with biologically propelled designs and arrange in layers architecture.
In their work, the IEEE 13 Node Test Feeder [20] model system is utilized to identify any power theft.
It includes 24 data sets that model the power stream of every bus in 24 hours, this could acquire users'
utilization consistently. ANN can recognize power theft using three techniques: (1) separates the entire
system into smaller section, which the power theft can be distinguished according to the plot, (2) splits
the system into three stages for each stage of the system is treated as one research object, it utilized recursive
Bayesian approach to detect load stage in distribution system and (3) like (1) consider the system as one
research object. This work aims to prevent overfitting ANNs in the future with an exponential volume
of train data and increase the efficiency in selecting training data from history but no specific calculation was
presented. A better criterion to detect will be developed to identify dishonest users.
Authors in [21] have classified data falsification attacks into four types: additive, deductive,
camouflage and conflict. In additive attack, there is a possibility which an adversary may tamper the smart
meter data by adding up the values which may result in increased electricity bills which can damage
customer confidence to power utility due to higher bills compare to other competitors. But if the utility
companies participate in the demand response, these companies may experience losses in revenue due to
the additive attack for unnecessary payment paid to the customer for caused peak shifts. While in deductive
attack is referred to loss of income for electricity utility corporations. Cover-up or camouflage happens where
the attacker will split the compromised meter into two groups with the same numbers that concurrently adopt
an additive and deductive mode, correspondingly. The attack will stay undetected because of the overall
input and output of the smart meters and the total demand and the whole usage of the power consumption
may remain unchanged. Lastly conflict attack is a scenario that is reprensted as random attack of existence
of additive and deductive attacks both in prevention of both attack from being detected. This attack is mixed
type of attack with the unequal margin of falsification for each hidden attack type. In their study, Pecan
Street [22] dataset is used to detect the data falsification attack using statistical technique. Even though
the end result effiecienly shows an high detection rate of 98% with 7% of false positive but how the attack
dataset generated was not presented and was based on intuitive of the reserachers.
Authors in [23] have suggested an artificial intelligence algorithm to detect electricity theft by using
the machine learning feature space. This novel framework called electricity theft detection (ETD) which
include a smart algorithm like k-nearest neighbour and ETD to identify fake buyer from the real customer
using consumption behavior. ETD utilizes transformer meters and observes any anomalies in customer
consumption patterns in providing a good solution to detect power theft in terms of cost efficiency while
k-Nearest Neighbor makes the framework more robust with the retraining capability. The irish social science
data archive center (ISSDA) [18] is used as the real-time dataset to validate their approach where the result of
the accuracy for k-Nearest Neighbour is 88% while ETD is 93% which makes the ETD robust against any
non-malicious changes in consumption pattern. Authors in [24] have proposed the entropy-based electricity
theft detector (EBETD) approach to identify any energy theft activities. In their study, MATLAB is used to
detect energy theft activities by tracing the dynamic behavior of the data consumed by the end-users. Entropy
is used to measure the values between two random distributions when the attackers sends the tampered
electricity consumption reading toward the control center. The entropy will compare the reading with the
predefined threshold numbers that were preset using previous historical power consumption variation using
entropy calculation. The control center will conclude that attack has been launch by the adversary. If when
compared with each sample during the runtime at each time step produce an entropy value that is more than
the predefined threshold value. The irish social science data archive center is used and 5 types of attack for
energy theft were demonstrated:

 ISSN: 2302-9285
270
a. A1, attacker changed the sequence number of the power consumption by replacing the first number with
the last and submitted it to the control center
b. A2, adversary used the mean value from previous day and send it to control center
c. A3, the previous day's consumption of mean value of the
d. A4, multiplied all power consumption with random number selected
e. A5, multiplied using same random number on particular day
The result in their study shows that EBETD approach has detected the attack with high positive rate
(DR). The detection attacks of A1 and A2 are 81.24 and 98.92% DR. While attacks A3, A4, and A5
are 77.03, 100, and 94.22% DR. Lastly, the trust level will achieve high level. As in attack A4, the DR
is 100% and FPR is 23.13% when trust level is 90% so they set in their study as the trust level is 99%.
Jokar [25] proposed ROC Curve tools to be used in order to detect energy theft by observing any anomalies
in customer consumption trend. It shows a novel of consumption pattern-based energy theft detector
(CPBETD), that used consistency instance of normal users and example from the pattern of malicious
consumption. CPBETD is quite effective dealing with contamination attacks and non-malicious changes in
consumption trends. The support vector machine (SVM) technique is used to predict the target value of
the test data. The data vector is mapped to a higher dimensional space where information classes
are progressively recognizable. At that point, an isolating hyper-plane with the most extreme edge to
the nearest information focuses on each class is found. The irish smart energy trial [18] dataset is used in the
study and its record half-hourly interval electricity usage reports for more than 5000 businesses
and homes from 2009 to 2010. The overall performance of CPBETD algorithm used in this study,
the accuracy obtains 95% while the detection rate 94% if the lower the value of Bayesian detection rate
(BDR) while considering the tampering rate 0.72%.
Cody [26] in his work has proposed decision tree learning for distinguishing fake activity in
fine-grained power consumption information. In their study, WEKA is used as a data mining tool to
demonstrate a successful application of the proposed technique by using the root mean squared error [16]
which is a statistical technique to calculate the variance between anticipated values and real values. M5P will
present the normal user power consumption behavior which can predict future power consumption and trace
any malicious activity. The outcomes exhibited the M5P decision tree learning algorithm can be used in
tracing energy fraud. The dataset will be segmented into multiple branches according to the decision rules
base on the input attributes and outputs. The irish social science data archive center [18] dataset is utilized in
this work used to present the efficiency of their methodology. They aimed at improving the detection and
prediction based on this model in their future study. The feature extraction algorithm can be considered to
enhance the model in pre-processing methods with the ability to remove irrelevant and redundant features.
Then correlation with other machine learning methodologies can be led in recognizing the positive
and negative parts, for more robust fraud detection system.
In [27], ANNs have been recommended to identify fraud in energy consumption data. During the
evaluation, they used WEKA as a machine learning tool and indicate fraudulent activities by detecting
anomalies from each of the end user consumption behaviour. To calculated the variation between expected
and real meter reading in the social science data archive center (ISSDA) dataset [18], root mean squared error
(RMSE) [28] is used. A confusion matrix was then developed to measure the level of effectiveness of the
energy consumption behaviour (ECB) in ANN, to indicate the true positives (TP), true negatives (TN), false
positive (FP) and false negative (FN). In his approach, if TP with 93.75% reflect that the ratio of fraudulent
activities classified accurately as data theft. While TN of 75.00% classified as normal data with normal
activities. The FP of 25.00% is considered as fraud. Lastly, FN corresponds to the numbers of fraudulent
activities that classified as normal behavior. TP highest value will indicates that the ECB's neural network manage
to detect any anomalies of malicious behaviors in the dataset. While the FP of 25.00% shows some of the normal
activity in NN was labeled as fraud. Although the result from experiment in this work was very positive but how
the attack can changed the user dataset and what changes does attacker made is not presented.
The authors in [29] suggested a hybrid clustering technique to identify the vulnerability node cluster
against the false data injection (FDI) attack. In their study, MATLAB is used to identify FDI attack on
the AMI. The level of FDI is then classified based on the vulnerabilities poses by every node in AMI,
the associated relationship of various nodes is obtained and clustered using enhanced constriction factor
particle swarm optimization (CF-PSO) will categorize each node into the least, moderate and most vulnerable
group. Using lengthy experiments utilizing two test systems, have shown that group with higher likelihood is
more prone to undermining the operation of the system. The UCI machine learning repository [30] dataset is used
which contains 58,000 user data including numerical attributes in a total of 9. The dataset is used to improves
the clustering approach by evaluating the performance of the k-means and CF-PSO. They concluded in their
work that verifying the same characteristics of nodes under FDI attacks is important in detecting this attack.
In this work, they managed to present how a node in AMI can be more likely to be attacked in AMI. In their

271
future work, where the initial work is already under preparation, they suggested the protection and detection
strategy in identifying the most vulnerable nodes in AMI. Summary of past literature for anomaly detection
techniques (ADTs) in AMI shown in Table 1.
Table 1. Summary of past literature for ADTs in AMI
Publication
Parameter
Ref. Focus Dataset Tools Technique Result
[5]
A practical feature-
engineering framework
for electricity theft
detection in smart grids
Irish CBT
AUC
performance
FMM clustering
algorithm
Accuracy 0.684 -
0.811. Computation
time: 10.224 min -
82.113 min
Accuracy &
Computation
Time
[9]
Anomaly Detection with
Time Series Forecasting
Data is a use
case e.g.,
revenue, traffic)
and it is at a day
level with 12
metrics
Python,
Keras &
Auto Arima
Time series
forecasting
Time series
forecasting helps to
prepare for future
needs by estimating
anomalies with the
current data.
Computation
Time
[17]
Electricity theft detection
with the random tuning
of hyperparameters
Irish Smart
Energy
Tensorflow
(Python)
NN
Detection rate - 93%
False acceptance -
5%
Accuracy
[19]
Energy theft detection
via ANN Not Stated
Radial
Distribution
Test
Feeders
ANNs Future work Accuracy
[12]
Behavioral Based Trust
Metrics and the Smart
Grid
Real-world
cyber Security
Gateway Syslog
MCM,
BEA, and
FCEA
processors
Machine and
statistical learning
algorithms
The methods were
found to be accurate
with low false
positive and negative
rates
Trust
[14]
Improving User Trust on
Deep NN based Intrusion
Detection Systems
NSL-KDD
Layer-wise
relevance
propagation
Binary classification
and multiclass
classification
Binary classification
(A ~97%; intrusions
detected).
Multi-class
Classification (A
~94%; DoS attack
detected at A ~98%)
Trust
[21]
Data falsification on
smart meter consumption
Pecan
Not
mentioned
Statistical
Detection rate - 98%
False alarm rate - 7%
Accuracy &
Trust
[23]
Detecting fraudulent
consumer from the
normal consumer based
on the consumer’s
consumption pattern
Irish Social
Science Data
Archive
Feature
Space
k-Nearest
Neighbour ETD
Algorithm
k-Nearest
Neighbours: A ~
88% ETD algorithm:
A ~ 93%
Accuracy
[24]
Detecting energy theft
using entropy
Irish Social
Science Data
Archive Center
MATLAB
R2014b
Entropy- based
detection
Detection rate:
A1=81.24 A2=98.92
A3=77.03 A4=100
A5=94.22
Trust=99%
Accuracy &
Trust T
[25]
Detecting energy theft by
monitoring abnormalities
in customer consumption
patterns
Irish Smart
Energy
ROC curve
Support Vector
Machine
Financial loss with A
~95%. detection rate
94%; tampering rate
0.72%
Accuracy
[13]
Adaptive Threshold
Selection for Trust-based
Detection Systems
Network flows
collected by a
European
Internet Service
Provider (ISP
Beta
probability
density
function
& ROC
Curve
Host-degree and
port-degree
experiments
The host degree’s
threshold for
STATIC-TPR is 5
and Port-degree’s
threshold value is the
same too. True
Positive Rate (TPR)
is above 70% over 14
days.
Trust
[26]
Fraud Detection in
Consumer Energy
Consumption
Irish Social
Science Data
Archive Center
WEKA
(Java)
Statistical Future work Accuracy

 ISSN: 2302-9285
272
Table 1. Summary of past literature for ADTs in AMI (Continue)
Publication
Parameter
Ref. Focus Dataset Tools Technique Result
[11]
Neighbor node trust-
based intrusion
detection system for
WSN
recommendation
database
MATLAB
Trust calculation of
neighboring node
The average
detection rate is 0.8
per 100 results.
Trust
[27]
Energy fraud
detection in energy
consumption data
Irish Social
Science Data
Archive Center
WEKA
(Java)
ANNs
A: Fraud FP
~25.00%; True
negative ~75.00%
True positive
~93.75% FN ~6.25%
Accuracy
[29]
Identification of
vulnerable node
clusters against FDI
attack
UCI Machine
Learning
Repository
MATLAB
Hybrid clustering
technique
Future work -
[10]
Energy-Theft
Detection Issues for
Advanced Metering
Infrastructure in
Smart Grid
Not stated Not stated
Attack tree-based
threat model and
categorization of
energy-theft
Current AMI energy-
theft detection
schemes might be
classification-based,
state estimation-
based, or game
theory-based.
Computation
Time
[31]
Computational
Complexity of
Anomaly Detection
Methods
DARPA1999 Not stated
Entropy-based
Multidimensional
Mahalanobis-
distance Method
(EMMM) and
CSDM method of χ2
value using multi
statistical variables
The total time
complexity of the
EMMM method is
O(n) for the n total
packets.
Computation
Time
4. CONCLUSION AND FUTURE WORKS
Various potential security issues related to AMI have been identified and an actual threat scenario
has been implemented confirming the vulnerability of AMI environments. Apparently, ADTs are going to
face challenges and difficulties to combat such rapidly emerging new threats. It is essential to identify
abnormal activities that can decrease the functional risks, unexpected downtime and to avoid unseen
problems of the AMI related components. Thus, detection of anomalies mechanisms needs to be continuously
upgraded and enhanced in order to increase the security level critical data in the AMI environment.
In summary, it is important to have a detection mechanism in place to ensure that data transmission in
the AMI infrastructure is secure and not easily manipulated. This review reveals the various suggestions
and implementation of ADTs. In evaluating the techniques, most existing studies focused on producing
accurate results. However, there are also a number of studies that focused on computation time and trust
issues. MATLAB is found as a popular tool of choice for demonstration and evaluation purposes.
This review may provide references in obtaining detailed information on various ADTs approaches
established in the AMI environment. Further improvement of the new techniques and procedures is required
to process the diverse data streams produced in the AMI environment in order to achieve an accurate
detection rate.
ACKNOWLEDGMENT
The authors would like to thank the Ministry of Higher Education Malaysia (MoHE) and Universiti
Tenaga Nasional (UNITEN) for funding this study under the Fundamental Research Grant Scheme (FRGS)
of Grant No. FRGS/1/2017/ICT03/UNITEN/03/1.
REFERENCES
[1] Edison and A. MeterComittees, "Smart Meters and Smart Meter Systems: A Metering Industry Perspective,"
An EEI-AEIC-UTC White Paper, pp. 1-35, 2011.
[2] Z.-A. Ibrahim, F. A. Rahim, R. Ismail, and A. A. Bakar, "A Review of Big Data Digital Forensic Analysis
in Advanced Metering Infrastructure," Advanced Science Letters, vol. 24, no. 3, pp. 1603-1607, 2018.
[3] G. Muruti, F. A. Rahim and Z. bin Ibrahim, "A Survey on Anomalies Detection Techniques and Measurement
Methods," 2018 IEEE Conference on Application, Information and Network Security (AINS), Langkawi, Malaysia,
pp. 81-86, 2018.

273
[4] Paul Bolton, Sarah Barber, "Energy Smart Meters," House of Commons Library Brieﬁng Paper, no. 8119,pp. 1-26,
Oct 2019.
[5] R. Razavi, A. Gharipour, M. Fleury, and I. J. Akpan, "A practical feature-engineering framework for electricity
theft detection in smart grids," Applied Energy, vol. 238, pp. 481-494, March 2019.
[6] A. M. Al-Ghaili, H. Baskaran, Z. Ibrahim, F. A. Rahim and S. A. S. Hairi, "A Dynamical Behavior Measurement
Algorithm for Smart Meter Data: An Analytical Study," 2019 IEEE Conference on Application, Information and
Network Security (AINS), Pulau Pinang, Malaysia, pp. 66-70, 2019.
[7] K. Khalid, A. Mohamed, R. Mohamed, and H. Shareef, "Performance Comparison of Artificial Intelligence
Techniques for Non-intrusive Electrical Load Monitoring," Bulletin of Electrical Engineering and Informatics,
vol. 7, no. 2, pp. 143-152, 2018.
[8] M. Nawir, A. Amir, N. Yaakob, and O. B. Lynn, "Effective and efficient network anomaly detection system using
machine learning algorithm," Bulletin of Electrical Engineering and Informatics, vol. 8, no. 1, pp. 46-51, 2019.
[9] A. Krishnan. "Anomaly Detection with Time Series Forecasting," M toward data science, [Online]. Available:
https://towardsdatascience.com/anomaly-detection-with-time-series-forecasting-c34c6d04b24a, 2019.
[10] R. Jiang, R. Lu, Y. Wang, J. Luo, C. Shen and X. Shen, "Energy-theft detection issues for advanced metering
infrastructure in smart grid," in Tsinghua Science and Technology, vol. 19, no. 2, pp. 105-120, April 2014.
[11] S. M. Sajjad, S. H. Bouk, and M. Yousaf, "Neighbor Node Trust based Intrusion Detection System for WSN,"
Procedia Computer Science, vol. 63, pp. 183-188, 2015.
[12] J. Obert, A. Chavez and J. Johnson, "Behavioral Based Trust Metrics and the Smart Grid," 2018 17th IEEE
International Conference On Trust, Security And Privacy In Computing And Communications/ 12th IEEE
International Conference On Big Data Science And Engineering (TrustCom/BigDataSE), New York, NY,
pp. 1490-1493, 2018.
[13] Y. Chae, N. Katenka and L. Dipippo, "Adaptive Threshold Selection for Trust-Based Detection Systems," 2016
IEEE 16th International Conference on Data Mining Workshops (ICDMW), Barcelona, pp. 281-287, 2016.
[14] K. Amarasinghe and M. Manic, "Improving User Trust on Deep Neural Networks Based Intrusion Detection
Systems," IECON 2018 - 44th Annual Conference of the IEEE Industrial Electronics Society, Washington, DC,
pp. 3262-3268, 2018.
[15] Y. Jin and S. Min, "Stadam: A new SLA trust model based on anomaly detection and multi-cloud," 2016 First
IEEE International Conference on Computer Communication and the Internet (ICCCI), Wuhan, pp. 396-399, 2016.
[16] G. Martin, "Electricity smart metering customer behaviour trials findings report, Technical report," CER
Commission for Energy REgulation, pp. 1-146, 2011.
[17] M. Nabil, M. Ismail, M. Mahmoud, M. Shahin, K. Qaraqe and E. Serpedin, "Deep Recurrent Electricity Theft
Detection in AMI Networks with Random Tuning of Hyper-parameters," 2018 24th International Conference on
Pattern Recognition (ICPR), Beijing, pp. 740-745, 2018.
[18] J. McBride, "What is the Irish Social Science Data Archive?," Journal Irish Political Studies, vol. 17, no. sup 1,
pp. 1-3, 200.
[19] H. Huang, S. Liu and K. Davis, "Energy Theft Detection Via Artificial Neural Networks," 2018 IEEE PES
Innovative Smart Grid Technologies Conference Europe (ISGT-Europe), Sarajevo, pp. 1-6, 2018.
[20] W. H. Kersting, "Radial distribution test feeders," in IEEE Transactions on Power Systems, vol. 6, no. 3,
pp. 975-985, Aug 1991.
[21] S. Bhattacharjee, A. Thakur, S. Silvestri, and S. K. Das, "Statistical Security Incident Forensics against Data
Falsification in Smart Grid Advanced Metering Infrastructure," Proceedings of the Seventh ACM on Conference on
Data and Application Security and Privacy, Scottsdale, Arizona, USA, pp. 35-45, 2017.
[22] N. None, "Pecan Street Grid Demonstration Program. Final technology performance report,"; Pecan Street Project,
Inc., Austin, TX (United States), 2015. [Online]. Available: https://www.osti.gov/servlets/purl/1172297.
[23] R. Sowndarya and P. Latha, "An Artificial Intelligent Algorithm for Electricity Theft Detection in AMI,"
International Journal of Engineering Science and Computing, vol. 7, no. 3, pp. 5222-5227, 2017.
[24] S. K. Singh, R. Bose and A. Joshi, "Entropy-based electricity theft detection in AMI network," in IET Cyber-
Physical Systems: Theory & Applications, vol. 3, no. 2, pp. 99-105, 2018.
[25] P. Jokar, N. Arianpoo and V. C. M. Leung, "Electricity Theft Detection in AMI Using Customers’ Consumption
Patterns," in IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 216-226, Jan 2016.
[26] C. Cody, V. Ford and A. Siraj, "Decision Tree Learning for Fraud Detection in Consumer Energy Consumption,"
2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA), Miami, FL,
pp. 1175-1179, 2015.
[27] V. Ford, A. Siraj and W. Eberle, "Smart grid energy fraud detection using artificial neural networks," 2014 IEEE
Symposium on Computational Intelligence Applications in Smart Grid (CIASG), Orlando, FL, pp. 1-6, 2014.
[28] M. P. Anderson, M. P. Anderson, William W. Woessner, and R. J. Hunt Eds, "Applied Groundwater Modeling:
Simulation of Flow and Adjective Transport,". San Diego: Academic Press, 2015.
[29] A. Anwar, A. N. Mahmood, and Z. Tari, "Identification of vulnerable node clusters against false data injection
attack in an AMI based Smart Grid," Information Systems, vol. 53, pp. 201-212, 2015.
[30] Anand Sarwate, "UCI Machine Learning Repository,"An Ergodic Walk, 2013.
[31] S. Oshima and T. Nakashima, "Computational Complexity of Anomaly Detection Methods," 2012 Seventh
International Conference on Broadband, Wireless Computing, Communication and Applications, Victoria, BC,
pp. 644-649, 2012.

A Review of anomaly detection techniques in advanced metering infrastructure

More Related Content

What's hot (20)

Similar to A Review of anomaly detection techniques in advanced metering infrastructure (20)

More from journalBEEI (20)

Recently uploaded (20)

A Review of anomaly detection techniques in advanced metering infrastructure