In the world, the issue of electricity consumption has become a serious challenge. The importance of the energy issue has made careful monitoring of electricity consumption patterns to ensure stability, identify broken devices, combat energy theft, and generally detect anomalies an essential issue in the energy field. There are various methods for designing anomaly detection algorithms. However, among them, methods based on machine learning, especially deep models, have received much attention from researchers. However, some existing methods face limitations in accurately identifying electricity consumption anomalies. These limitations can be partly due to the limited availability of labeled data representative of diverse consumption patterns, as well as the difficulty in extracting effective features to describe complex and nonlinear energy consumption behaviors. However, some existing methods face limitations in accurately identifying electricity consumption anomalies. These limitations can be partly due to the limited availability of labeled data representative of diverse consumption patterns, as well as the difficulty in extracting effective features to describe complex and nonlinear energy consumption behaviors. To fill these gaps, this research presents an anomaly detection method in energy consumption data based on two-dimensional transforms and Convolutional Neural network (CNN) hybrid networks. Of course, this research mainly focuses on improving the representation of features and increasing the classification accuracy and does not directly address the problem of data scarcity. In the proposed method, after preprocessing, the existing data are converted into two-dimensional data using Continuous Wavelet Transform (CWT) and Recurrence Plot (RP) transformations in the next step. Then, two dedicated CNN architectures are designed for two types of two-dimensional output data, CWT and RP. The task of each of these networks is to classify the input two-dimensional data into one of the two main classes of normal and abnormal. Finally, there is a result combination block that converts the results of the two networks into a final result as a weighted average. The simulation was investigated using the LEAD dataset containing about 1.7 million hourly readings of commercial building energy consumption. The samples were labeled at the consumer level into two classes: normal and abnormal, which after data preparation included 1,522,060 normal samples and 227,434 abnormal samples; the data were also divided into training, validation, and test sets in the ratio of 70%, 15%, and 15%. The performance of the method was evaluated using the confusion matrix and the criteria of Sensitivity, Specificity, Accuracy, Precision, and F1-Score, and the values were 93.88%, 95.23%, 96.17%, 74.63%, and 83.15%, respectively, which shows the efficiency of the proposed method in identifying abnormal electricity consumption patterns.
