Electricity theft (ET) is a major problem in developing countries. It a�ects the
economy that causes revenue loss. It also decreases the reliability and stability
of electricity utilities. Due to these losses, the quality of supply e�ects and tari
� imposed on legitimate consumers. ET is an essential part of Non-technical
loss (NTL) and it is challenging for electricity utilities to �nd the responsible people.
Several methodologies have developed to identify ET behaviors automatically.
However, these approaches mainly assess records of consumers' electricity usage,
may prove inadequate in detecting ET due to a variety of theft attacks and irregularity
of consumers' behavior. Moreover, some important challenges are needed
to be addressed. (i) The number of normal consumers has been wrongly identi�ed
as fraudulent. This leads to high False-positive rate (FPR). After the detection
of theft, on-site inspection is needed to validate the detected person, either is it
fraudulent or not and it is costly. (ii) The imbalanced nature of datasets which
negatively a�ect on the model's performance. (iii) The problem of over�tting
and generalization error is often faced in deep learning models, predicts unseen
data inaccurately. So, the motivation for this work to detect illegal consumers
accurately.
We have proposed four Arti�cial intelligence (AI) models in this thesis. In system
model 1, we have proposed Enhanced arti�cial neural network blocks with skip
connections (EANNBS). It makes training easier, reduces over�tting, FPR and
generalization error, as well as execution time. Temporal convolutional network
with enhanced multi-layer perceptron (TCN-EMLP) is proposed in system model
2. It analyzes the sequential data based on daily electricity-usage records, obtained
from smart meters. At the same time, EMLP integrates the non-sequential
auxiliary data, such as data related to electrical connection type, property area,
electrical appliances usage, etc. System model 3 based on Residual network (RN)
that is used to automate feature extraction while three tree-based classi�ers such
as Decision tree (DT), Random forest (RF) and Adaptive boosting (AdaBoost)
are trained on the obtained features for classi�cation. Hyperparameter tuning
toolkit is presented in this system model, named as Hyperactive optimization
toolkit. Bayesian is used as an optimizer in this toolkit that aims to simplify
the tuning process of DT, RF and AdaBoost. In system model 4, input is forwarded
to three di�erent and well-known Machine learning (ML) techniques, i.e.,
Support vector machine (SVM), as an input. At this stage, a meta-heuristic algorithm
named Simulated annealing (SA) is employed to acquire optimal values
for ML models' hyperparameters. Finally, ML models' outputs are used as features
for meta-classi�ers to achieve �nal classi�cation with Light Gradient boosting
machine (LGBM) and Multi-layer perceptron (MLP). Furthermore, Pakistan residential
electricity consumption dataset (PRECON1), State grid corporation of
china (SGCC2) and Commission for energy regulation (CER3) datasets is used in
this thesis. SGCC dataset contains 9% fraudulent consumers, which are extremely
less than non-fraudulent consumers, due to the imbalance nature of data. Furthermore,
many classi�cation techniques have poor predictive class accuracy for the
positive class. These techniques mainly focus on minimizing the error rate while
ignoring the minority class. Many re-sampling techniques are used in literature to
adjust the class ratio; however, sometimes, these techniques remove the important
information that is necessary to learn the model and cause over�tting. By using
six previous theft attacks, we generate theft cases to mimic the real world theft
attacks in original data. We have proposed the combination of oversampling and
under-sampling techniques that is Near miss borderline synthetic minority oversampling
technique (NMB-SMOTE), Tomek link borderline synthetic minority
oversampling technique with support vector machine (TBSSVM) and Synthetic
minority oversampling technique with near miss (SMOTE-NM) to handle imbalanced
classi�cation problem. We have conducted a comprehensive experiment
using SGCC, CER and PRECON datasets. The performance of suggested model
is validated using di�erent performance metrics that are derived from Confusion
matrix (CM).