| Electricity theft not only causes serious damage to the economic benefits of power grid companies,disrupts the normal order of the electricity market,but also endangers power equipment and public safety.Due to the high cost of deploying electricity stealing detection devices for power grid companies,and facing complex and randomly fluctuating electricity load curves,the current detection algorithm is prone to misjudging normal users as electricity theft users;It greatly increases the detection difficulty,and the data imbalance seriously affects the detection accuracy,and most of the single detection methods currently used cannot cope with all different application scenarios.In addition,in order to achieve the goal of "carbon peaking and carbon neutrality",the vigorous development of new energy power generation,mainly photovoltaic power generation,has spawned the behavior of electricity theft that defrauded high subsidies by inflating electricity,gradually harming national interests and affecting power grid security.This type of electricity theft has not received enough attention,and there are few related studies.In this context,this paper takes the data of various scenarios provided by power grid companies as the research object,focusing on cross-industry hybrid automatic detection of electricity consumption,targeted detection of multi-source data electricity theft,low-cost detection of a small amount of tag data,and distributed photovoltaic theft In four aspects,such as low false detection rate detection,we strive to enrich the related research on detection methods in different application scenarios.The main research work and innovation points of the paper are as follows:(1)Aiming at the problem of complex power consumption scenarios and high randomness of power consumption resulting in mixed load curves,and it is difficult for a single model to accurately identify abnormal power consumption,a multi-model fusion integration based on stacking structure is proposed under the condition of only using power consumption data.The learning algorithm detects the behavior of electricity stealing,and uses different strong models to study the data from different angles.Firstly,the basic principle of the ensemble learning method is explored,and the algorithm mechanism of each individual learner of the base model is analyzed.Secondly,the stacking structure is deduced step by step,including the feature extraction process,the selection and analysis of the base model and meta model,and the selection of hyperparameters.Experience is difficult to optimally choose the disadvantage of the base model.Finally,the experimental verification is carried out using the dataset of State Grid Corporation of China,the model AUC value is0.98675.(2)Aiming at the problem that it is difficult to fully describe the characteristics of users’ electricity consumption behavior with a single electricity consumption data,this paper studies how to improve the recognition accuracy and reduce the misjudgment rate in the multi-data fusion scenario,and proposes an improved convolutional neural network based on channel attention network.Model-based electricity stealing behavior detection method.Firstly,a power theft evaluation index system including power consumption trend,line loss growth rate and terminal alarm multi-source data fusion is established under the metering automation system of the power grid company,so as to construct a user power consumption feature set.Secondly,the convolutional neural network is optimized based on the channel attention network,and the CNN(Convolutional Neural Network)model based on the channel attention network SENet(Squeeze and Excitation Networks)is established accordingly.loss optimizes the model loss function to solve the dataset imbalance problem,thereby improving detection accuracy.Finally,the experimental verification is carried out using the data set of China Southern Power Grid,,the model AUC value is as high as 0.999733.(3)In view of the high cost and difficulty of obtaining labeled data for power grid companies,and the unlabeled data obtained is difficult to train an effective electricity theft detection model,a generative adversarial network CT-GAN(Co-training Generative Adversarial Networks)semi-supervised electricity theft detection method.First,the principles and structures of generative adversarial networks and semi-supervised generative adversarial networks are explored.Secondly,the Wasserstein distance is proposed to replace the JS(Jensen-Shannon)divergence and KL(Kullback-Leibler)divergence distance to solve the problem of poor model training and poor quality of generated data caused by gradient disappearance and mode collapse in Generative Adversarial Networks problem,and built a multi-discriminator joint training model to avoid the problem of high distribution error of a single discriminator,while enhancing the ability of GAN to generate label sample data,and by expanding the label sample data set,to improve the model detection accuracy and generalization ability.Finally,the experimental verification is carried out using the actual dataset of the Irish Smart Energy Trail,when the label data contains 6.25%,the classification accuracy of the model is as high as 0.8142.(4)When the distributed photovoltaic power station is connected to the power grid on a large scale,there is a behavior of stealing electricity by cheating high subsidies by increasing the amount of electricity,and the blindness of on-site inspection of photovoltaic electricity stealing is high due to the inability to obtain the type of electricity stealing from users suspected of stealing electricity.Due to non-human factors such as equipment failure and rapid climate change,the detection algorithm has a high false detection rate.Under the condition that only the power generation data of photovoltaic power plants are used,a one-dimensional convolution based on Attention Mechanisms(AM)is proposed.1DCNN-AMBL electricity stealing detection method combined with neural network(1D Convolutional Neural Network,1DCNN)and Bi-directional Long Short Term Memory(Bi-LSTM).First,the power generation and grid-connection mechanism of typical distributed photovoltaic power plants and the output correlation analysis of each power generation unit are studied.Secondly,a distributed photovoltaic power generation electricity stealing detection model based on 1DCNN-AMBL is constructed,and on this basis,an objective function is constructed to optimize the false detection rate of the model,which ensures the model detection rate and reduces the false detection rate.Finally,the experimental verification is carried out using the power generation data set of the actual distributed photovoltaic power station in a certain area provided by Ausgrid Power Grid Company on the east coast of Australia,the false positive rate of the model is only 0.0061 and the AUC value is 0.9846.(5)In view of the fact that each provincial power grid company has not established a unified electricity theft detection platform,and the data types provided by the metering automation systems of each provincial power grid company are different,and the existing intelligent detection methods cannot deal with the electricity theft problem in all scenarios,the development of The hardware platform for the reproduction of electrical behavior and the Lab VIEW software platform for the host computer for self-adaptive diagnosis of electricity theft.Firstly,based on the research on the mechanism of electricity theft and the actual meter types of power grid companies,different methods of electricity theft reproduction are studied to obtain electricity theft data on the developed electricity theft reproduction hardware platform.Secondly,the acquired data has been embedded into the Lab VIEW software platform to identify the aforementioned self-adaptive diagnosis algorithm for electricity theft,and display the detection results in real time,so as to realize the design and detection of the software and hardware platform integrating the generation and discrimination of electricity theft data.Finally,the data sets obtained from the electricity theft recurrence platform are used to verify the aforementioned electricity theft detection methods.The research focus of this work is to study how to accurately,targeted and effectively detect abnormal electricity users and electricity theft behavior,and at the same time solve how to reduce the detection cost and the probability of false detection events,and enrich the detection methods of power grid companies for different application scenarios.Provide accurate and effective detection methods for the current smart grid to discover abnormal electricity consumption behaviors and inspect users who steal electricity. |
PDF Full Text Request