**1. Introduction**

Electricity theft can be defined as the behavior of illegally altering an electric energy meter to avoid billing. This illegal behavior not only severely disrupts the normal utilization of electricity, but also causes huge economic losses to power systems. At the same time, the unauthorized modification of lines or meters easily leads to accidents such as power failures and fires, and poses a serious threat to the safety of the relevant power system [1,2]. According to the research released by an intelligence firm northeast group, llc in January 2017, electricity theft and other non-technical losses have rendered over \$96 billion in losses per year globally [3]. State Grid Hunan Electric Power Company, China reported that nearly 40% of electrical fires and 28% of electric shock casualties are caused by electricity theft [4]. Therefore, it is necessary to develop e ffective techniques for electricity theft detection and ensure the security and economic operation of power system.

The electricity theft detection technologies can be divided into three categories: the network-oriented method, the data-oriented method and a hybrid-oriented method that mixes the previous two methods [5]. Network-oriented and hybrid-oriented approaches usually require the network topology [6,7] and even the installation of additional devices [8]. It is di fficult to implement these approaches widely, because the network topology may be unattainable due to security concerns and

the installation of addition devices is costly. Data-oriented approaches only focus on the data provided by smart meters and have no requirements of the network topology or additional devices, which helps with improving the cost-e ffectiveness for suspected electricity theft judgment and detection. Therefore, data-oriented approaches have been widely applied to electricity theft detection in recent years [9,10].

At present, there are two typical data-oriented methods to detect electricity theft: support vector machines (SVM) and neural networks. In [11], they proposed a SVM-based approach that uses customer consumption data to expose abnormal behavior and identify suspected thieves. The authors in [12,13] combined SVM and a fuzzy inference system to detect electricity theft. In [14], a comprehensive top-down scheme based on decision trees and SVM was proposed. The two-level data processing and analysis approach can detect and locate electricity theft at every level in power transmission and distribution. The authors in [15] proposed an ensemble approach combining the adaptive boosting algorithm and SVM.

More and more researchers are utilizing neural networks to detect electricity theft due to their effectiveness. In [16], a long short-term memory (LSTM) and bat-based random under-sampling boosting (RUSBoost) approach is proposed. The LSTM and bat-based RUSBoost are applied to detect abnormal patterns and optimize parameters, respectively. In [17], a method based on the wide and deep convolutional neural network (CNN) model is proposed. The deep CNN component can identify the periodicity of electricity consumption and the wide component can capture the global characteristics of electricity consumption data. The authors in [18] combined CNN and LSTM to detect electricity theft from the power consumption signature in time-series data. In [10], an end-to-end hybrid neural network is proposed, which can analyze daily energy consumption data and non-sequential data, such as geographic information.

The above methods have paved the way for building the structures of networks and dedicate to improving electricity theft detection's accuracy. However, they mainly focus on the daily or weekly electricity consumption patterns. As a result, if these methods are applied to hourly or more frequent electricity consumption data, their accuracy will decrease. This is because they fail to capture the intraday electricity consumption pattern, for example, the correlation between the electricity consumption at the same time on di fferent days. In practice, some illegal users commit electricity theft for part of the day. Specifically, those who have precedent technology and large electricity demands (such as industrial electricity thieves) prefer to commit the crime during some specific hours after considering electricity prices, monitoring periods and the risk of being caught comprehensively. Meanwhile, new kinds of attacks such as interception communication and false data injection [19,20] make it easier to commit such crime. In a case of electricity theft caught by State Grid Shandong Electric Power Company, China, the illegal user had normal daily electricity consumption. However, he confused the metering time of his smart meter to avoid peak electricity tari ffs [21]. In this way, his abnormal electricity consumption pattern can only be reflected in the intraday data. In paper [22], several possible attack models are proposed to confuse the metering time and commit electricity theft targeting time-based pricing. Therefore, it is necessary to construct an electricity theft detection scheme that can not only capture the daily features but the intraday features.

In order to better extract the periodical features from days and more frequent time periods, we utilize a two-dimensional grid structure for the raw input data in this paper. Based on the data structure, we propose a text convolutional neural network (TextCNN) to detect electricity theft. The main contributions of the proposed model include:


TextCNN. To test the performance, we implemented extensive experiments on the residential and industrial datasets from a province in China and the public Irish residential dataset.

(3) We propose the data augmentation method to expand the training data in view of the shortage of electricity theft samples. Experimental analysis indicates that the method can improve the detection accuracy effectively with a proper augmentation process.

The remainder of this paper is organized as follows. The methodologies of data construction and CNN construction for electricity consumption data are described in Section 2. Section 3 proposes the neural network structure based on TextCNN and the data augmentation method. Then, the details of experimental datasets and methods used for comparison and metrics are given in Section 4. Section 5 presents the performance and superiority of the proposed model, analyzes the parameters and discuss the effectiveness of the data augmentation method. Finally, Section 6 concludes this paper.
