CNN-AdaBoost based hybrid model for electricity theft detection in smart grid

As the use of deep learning models is increased in smart grid systems, especially in load forecasting, supply-demand response, vulnerability detection, and finding abnormal behavior of the customers, it becomes necessary to find out the models' flaws and increase their classification accuracy. While designing robust and secure machine learning (ML) algorithms for real-world applications, artificial vulnerabilities, data imbalance problems and choosing the right algorithm play a major role. The paper proposed a hybrid method that deals with different issues like the curse of dimensionality, data imbalance problem and also study existing models which give low theft detection rate. By considering these problems and flaws in existing theft detection models, this hybrid model is developed. The model being proposed is divided into various phases. The first module is used to handle data preprocessing which comprises handling issues like outliers, missing values, and data imbalance. The second module contains a convolutional neural network (CNN), is takes the processed data from the first module. CNN pulls the essential features from the processed data and the new dataset is transferred to adaptive boosting also called AdaBoost algorithm which performs the classification of legitimate users and fraud users. Based on the results obtained from the SGCC dataset, it has been verified that the hybrid framework proposed is capable of generating highly accurate and reliable classification performance when compared to other established and advanced machine learning models. The proposed system can be used practically to detect power theft in industrial and commercial applications.