Abstract:
Electricity theft is a major problem facing many countries around the world. Its adverse
effects include loss in revenue for power distribution companies and government
economy, the distribution quality of electricity, increased generation load and high electricity
cost which affects honest consumers as well. With the advancement in smart
meters in electricity infrastructure, massive data is generated that can be analyzed for
electricity consumers consumption patterns. By the help of these consumption patterns
of consumers several machine learning and deep learning models and techniques
are built to detect fraudulent consumers. By the help of the detection of the theft location
and fraud consumers energy distribution companies can impose fines on these
consumers that help them to reduce revenue losses. In this research a combination of
CNN and STN based model is proposed to detect electricity theft. STN is used in many
image classification problems along with CNN to enhance the performance of the models.
STN along with CNN is used for the first time in electricity theft detection. STN
is used to translate, rotate and scale the the original input. The dataset used in this
research is real customers electricity usage data publicly provided by the State Grid
Corporation of China (SGCC). SGCC dataset has missing values, and also has class
imbalance problems. The number of theft users is significantly lower than the honest
consumers, which is addressed by using Synthetic Minority Oversampling Technique
(SMOTE). Proposed model is compared with state of the art machine learning and deep
learning models and techniques and results show that the proposed model can identify
theft and normal users with greater accuracy.
