-
is the detection of fraudulent attempts which are made for the illegal
reduction of bill payments. In the case of electricity, these attempts are made by
reversing the meters, by-passing or slowing down the meters or inaccurate
readings. The detection of theft attempts in power industry is termed as
Non-Technical Loss (NTL) detection. With the increasing demand for
electricity, the occurrences of NTL have been reported globally including India,
Pakistan, Brazil and China etc. In this paper, we first describe the use of the
synthesized and the real datasets in NTL detection. Then, we highlight an
interesting characteristic of class imbalance that is exhibited in the datasets
used for NTL detection. Moreover, we identify the fruitful areas in NTL
detection where the research community has been working on. Lastly, we
discuss the need for a relative comparison of the classical machine learning
and deep learning over a benchmark dataset for NTL detection.
1. Introduction
Recently, an increasing interest has been observed in recognizing the consumption
patterns of the consumers of electricity, gas and water supplies [
25
]. One of the main
objectives of this activity is to identify and forecast the potential theft attempts
in order to have reduced bills. This illegal theft attempt has dented the economies
of many countries causing a loss of billions of dollars. This includes China [
14
],
Pakistan [
10
], India, Brazil [
18
], etc.
Non-Technical Loss (NTL) detection in electric power industry is a term used
for the detection of faulty meters or illegal usage of electric units. Losses are bore
by the electric supply companies on the account of faulty meters that record fewer
units as compared to the consumed electricity. On the other hand, this practice
can also be intentional in order to get the electricity bill reduced by a substantial
margin. For both cases, the supplier companies look for a solution which can
identify them the faulty meters or potential theft instances.
One of the important characteristics of the relevant datasets is that they
belong to the class imbalance problem. It is the problem where the dataset is biased
towards one class by its heavy representation while the other class is less
representative. Interestingly, the problem becomes more challenging when the focus is on
the true representation of the less representative class. Naturally, the number of
normal electric consumption in a neighborhood area is huge as compared to the
number of theft attempts. This gives a clear indication that the real dataset of
the consumption of electricity belongs to the class imbalance problem where the
number of negative class samples is huge as compared to the number of positive
class samples. The techniques used in NTL detection should be able to balance
out the positive and the negative class samples before the dataset is used by the
machine learning algorithms [
9
].
One of the techniques used to identify the NTL is applying classical machine
learning algorithms to the datasets pertaining to the consumption of electricity.
This includes the use of Support Vector Machine (SVM), KNN, decision trees,
ensemble methods and neural networks [
8
]. Advances in deep learning has attracted
some researchers to test diferent variants of deep learning for NTL detection. For
e.g., the authors of [
3
] have used deep neural networks along with long short-term
memory network to identify the occurrences of NTL in a dataset pertaining to
the smart meters of a utility company in Spain. However, there is still a need
to compare the performances of the classical machine learning algorithms with
the diferent variants of deep learning architecture. In this paper, we focus on
elaborating the importance of a comparative study of the two paradigms for NTL
detection in a real dataset.
2. Synthesized Vs. Real Datasets
Two types of datasets are used in NTL detection. One of the types belongs to the
synthesized datasets which are randomly generated keeping in mind the
requirements of NTL detection. One of the benefits of using the synthesized datasets
is that they are easily accessible while on the other hand, they might miss the
potentially useful information which otherwise would have been handy in
detecting unlawful consumption activities. The other type is the real dataset which is
taken from a distribution company. An essential advantage of using such datasets
is analyzing the real and unique patterns of the consumption of electricity which
might be missing in the synthesized datasets. An associated drawback of using
real datasets is that it is hard to get an access of such datasets as distribution
companies generally avoid sharing the consumer’s information.
3. Class Imbalance: An aspect of NTL Detection
An interesting property associated with the datasets of electric distribution
companies is that they belong to the class imbalance problem. Considering two classes;
normal consumption and abnormal consumption; the class imbalance problem
arises when most of the consumption records belong to the normal consumption
class while few records belong to the abnormal consumption class. When the
objective is predicting the abnormal consumption class which is rarely represented
in the dataset, the success ratio of the classifiers might get deteriorated unless the
class imbalance ratio is properly dealt.
To balance out the number of representative records for both the classes,
undersampling the majority class can be used in the pre-processing stage of data
preparation. This is termed as synthetic minority over-sampling technique (SMOTE) [
30
].
The other way is over-sampling the minority class, i.e., duplicating or randomly
creating new synthesized records of the minority samples [
4
]. Both techniques have
been used in diferent problem domains belonging to the class imbalance. However,
few have tried these techniques in NTL detection.
4. Which Areas in NTL Detection are fruitful?
There is a general pattern followed in the pre-processing of the datasets and
application of the classifiers in order to achieve a significant success in NTL detection.
This pattern is shown in the Figure 1. The electricity distribution companies
provide raw data containing consumption records. Depending on the type of metering
infrastructure, the consumption data is either half hourly, hourly, daily, bi-monthly
or monthly records. The automated metering infrastructure (AMI) facilitates to
record hourly or daily consumption while the manual metering infrastructure uses
monthly manual readings done by the meter readers. Not all the records in
consumption profile are useful in performing analytics for NTL detection. Similarly,
not all the features are important for detecting NTL. For this, record selection and
feature selection is performed. Once the dataset is ready, it is scaled for
normalizing the values. The next step is training and testing of the classifiers. A range of
classifiers of diferent types is applied. The performance of the classifiers is then
measured by performance evaluation metrics.
Synthesized Vs.
Real Dataset
Raw Data
Post Processing
and NTL
Detection
Performance
Evaluation
Metrics
Monthly
Consumption
Records
Best Metrics
for NTL
Detection
Best Features
Data Munging/
Record Selection
Feature Selection
Training
and Testing
Classifiers
Best
Classifiers
Scaling for
Feature
Normalization
The researchers in this area are largely interested in three main modules. Some
of them try to find the best combination of features which are most suited for
NTL detection. For this, they use feature importance. In one of our previous
contributions, we propose the Incremental Feature Selection (IFS) algorithm that
selects the best combination of features responsible in detecting the occurrence of
NTL [
9
]. Some authors have taken a keen interest in finding the best individual
classifiers or a combination of classifiers for NTL detection. Others have tried to
ifnd the best type of the classifiers for NTL detection.
Some authors have focused on the performance evaluation module of the
classifiers. They have tried to find out which metrics are best suited to evaluate the
classifiers considering the specifications of the NTL detection problem. For
example, we have concluded in one of our previous works that recall should be given a
higher priority as compared to any other performance evaluation metric considering
the special characteristics of the NTL detection problem [
10
].
5. Machine Learning Vs. Deep Learning in NTL
Detection
Machine learning is mainly used for Non-Technical Loss (NTL) detection. NTL is
a typical classification task. Most of the researchers have used unsupervised,
supervised, semi-supervised, hybrid and network based methods. Some of the research
on NTL detection can be seen in [
2, 11–13, 15, 17, 19, 23, 24, 26–29, 31–36
]. Table 1
summarizes research work on NTL detection. Most researchers used electric supply
and distribution companies of specific countries as it is shown in Table 1. The most
common evaluation measures used in NTL detection are accuracy, precision, recall,
AUC and others as Table 1 shows. A framework to detect NTL is proposed in [
17
].
The characteristics of consumers were analyzed and data mining techniques were
used to recover electrical energy loss. Consumption data, extracted from a Spanish
power supply firm known as Endesa Distribution, was used by [
17
]. Association
rule mining was used to find groups of consumers responsible for non-technical loss
in the form of electricity theft. Hartmann et. al. [
13
] used contextual learning on a
dataset from Luxembourg electric distribution company. Accuracy, precision, recall
and F1 measures were used as evaluation measures. The attributes of consumption
were investigated in [
26
] for NTL detection in two firms. For this purpose,
hierarchical clustering, Benford curve and Multi-Dimensional Scaling (MDS) were used
on two datasets based on two Colombian companies. The study was evaluated using
ROC. Local Outlier Factor (LOF) based on the DBSCAN clustering algorithm was
used in the research work of [
28
]. The technique was evaluated using the Silhouette
coeficient and Davies Bouldin index. The research work of [
34
] used a synthetic
dataset and observed unusual profiles of electricity consumers by proposing a
distance matrix. Area Under ROC Curve (AUC), F-1 measure and accuracy was
used as evaluation measure. The results showed the efectiveness of the proposed
technique after comparing it with DBSCAN, GMM and kNN. Yeckle and Tang
[
33
] carried experiments on Irish dataset. Diferent outlier detection methods are
used to detect NTL. The performance is measured by AUC. The research work of
Zheng et. al. [
36
] explored deep convolutional neural networks (CNN) on a dataset
based Chinese electricity company. The result from deep convolutional neural
networks is compared with SVM, random forest, logistic regression and TSR (Three
Sigma Rule). It was observed that deep CNN outperformed the above mentioned
classifiers. Another research work of [
32
] used 4000 Irish household data records
collected from smart metering. Gustafson-Kessel clustering algorithm was used in
[
32
] to discriminate non-technical loss. The rate of true positive was observed as
63.6% whereas false positive was 24.3%.
A dataset collected from a Spanish electrical supply firm was analyzed in the
Evaluation
Accuracy,
precision,
recall,
measure
f1
Mean Index
Adequacy
and
execution time
(MIA)
ROC
Silhouette
coeficient,
Davies
Bouldin
index
Area Under
ROC Curve
(AUC),
accuracy and
F1 measures
AUC
AUC
True positive and false positive rate
Accuracy
Accuracy,
detection
rate,
precision, true
negative rate
research of [
11
]. Artificial neural network, regression and Self Organizing Map
(SOM) are used and as a result it was claimed that the accuracy has improved.
Statistical learning methods were used in [
23
] to detect NTL in South Africa. Three
classifiers namely Support Vector Machine (SVM), kNN (k Nearest Neighbour) and
naïve Bayes were used. The classification models were evaluated using accuracy,
detection rate, precision and true negative rate.
Less attention has been paid to deep neural networks and network science based
algorithms to detect NTL. From the literature review presented, it can be observed
that classical machine learning is the major streamline in solving the problem
of NTL detection. However, there is still a need of a thorough testing of diferent
variants of deep learning in a real dataset for NTL detection. Community detection
techniques have a key role in computer science, sociology, biology, physics,
economics, engineering, marketing, ecology, political sciences and many other
ifelds [
7
]. Community detection is used for inferring useful information from
complex networks. Complex networks are getting attention from researchers of
diferent domains. A network such as biological, information, technological,
social and other can be modelled as graphs [
6
]. Some well known community
detection techniques such as Greedy Modularity Maximization [
5, 16, 20
],
GirvanNewman algorithm [
21
], Louvain algorithm [
1
] and k-clique percolation [
22
] can also
be used to identify interesting patterns in NTL.