Abnormal Electricity Utilization Detection Based On Deep Learning And Few-shot Learning

Detection of abnormal electricity utilization is an important support for power grid security.With the construction of smart grid,power data has shown the key elements of large data,i.e.large volume and multiple types.Under the low density and noise interference of the mass data,the traditional detection modes are confronted with the difficulties of feature extraction,few labeled samples and unbalance of normal and abnormal samples.The detection efficiency is too low,and it is urgent to convert the direction of technology.In this paper,deep learning is applied to the detection of abnormal electricity utilization.With the automatically extracted features by the multi-layer nonlinear network,the network can effectively solve the difficulty of feature extraction,the subjective of threshold selection and so on.Two abnormal electricity utilization detection models which are based on semi-supervised learning and transfer learning are proposed to adapt to cases of few labeled samples and imbalanced samples.One is a semi-supervised model of AAE,which uses unlabeled data to optimize the auto-encoder network and uses labeled data to optimize the classifier.The combination of the two steps can not only preserve the local characteristics of the data,but also improve the generalization ability of the model.The other is a model based on CNN and transfer learning,which labels samples and creates two dimensional samples to solve the problem that the power data can not meet the requirements of the CNN training.The transfer learning method is used to adjust the model to enhance the learning of the uncertain samples.The experiments show the detection accuracies of two models are higher than the traditional detection methods.The semi-supervised model of AAE does not need any processing of the data,and the generalization ability is very adaptable.The model based on CNN and transfer learning can enhance the ability of features representation by few-shot learning and significantly increase the abnormal precision rate.The two methods can improve the hit rate of on-site recognition and reduce the cost of power companies.