An attention-based wide and deep CNN with dilated convolutions for detecting electricity theft considering imbalanced data

Abstract

• There is a serious sample imbalance in the data, which easily leads to a severe bias in the classification model, which in turn affects the classification accuracy. In order to solve the problem of sample imbalance, we proposed a method of using focal loss for model training to enhance the contribution of a small number of samples to model optimization. • Electricity consumption data is correlated in the time dimension. It is necessary to capture a larger receptive field to obtain global information and improve the model's accuracy. Keeping this concern in mind, we used dilated convolution to enlarge the receptive field without loss of information so that each convolution output contains a larger range of information, thereby capturing a larger receptive field and obtaining global information of the input data. • How to effectively integrate the feature expressions obtained by wide CNN and deep CNN is a key factor to improve the performance of the model. To address this problem, we adopt the channel-dimensional adaptive attention module to adaptively fuse the feature expressions obtained by wide CNN and deep CNN to improve the accuracy of model training. For the increasingly serious phenomenon of electricity theft, many researchers are trying to detect it. Traditional detection methods rely on physical inspection, which has low detection efficiency and high cost. However, the rapid application of advanced metering infrastructure (AMI) makes it possible to detect electricity theft via smart meters. In this work, we propose a hybrid improved wide and deep convolutional neural networks (CNN) method to address electricity theft detection on an imbalanced realistic dataset that presents a daily electricity consumption provided by State Grid Corporation of China. Our key contribution is the introduction of a channel-dimensional adaptive attention module concatenated with dilated convolutions. Moreover, we propose focal loss to solve the data imbalance problem. Experimental results demonstrate that compared to other state-of-the-art deep learning classifiers, our improved method has better performance (with 97.08% of MAP@100 and 83.61% of AUC), which verifies the effectiveness of the proposed method.