Anomaly Detection in Electricity Consumption Data: A CNN-based Approach for District-Level Analysis in Tehran, Iran

Detection of unauthorized electricity consumption in the smart distribution network is one of the main challenges in the implementation of this network and as a key part of the country’s energy distribution infrastructure. The reasons for any anomaly in consumption data are mainly divided into two categories: user error, and system error. In the case of user error, and in contrast to system responsibility, the responsibility for this problem lies with the user. Independent of the reasons for this issue, detection of this issue is carried out in two formats: real-time and based on long-term consumption data, each of which has its own advantages and disadvantages. In the case of reviewing long-term information, due to the presence of consumption data over a long period of time, hiding unauthorized consumption is associated with greater complexity. In this study, the consumption data of approximately 13,000 consumers in five of the Tehran electricity regions in the period one year (1401-1402) was used to train a deep learning-based model. Based on the collected data, and for each consumer, a binary label, corresponding to the authorized and unauthorized consumer, was considered. The definition of this label was performed by experts or based on field surveys. This model, which uses a convolutional neural network structure in its design, achieved a non-weighted average accuracy of 71.3% in the testing phase.