Research On Intelligent Detection Algorithms For Several Typical Abnormal Electricity Consumption

The accurate detection of abnormal electricity consumption(AEC)in power service marketing is vital to reducing non-technical losses and ensuring the safe management of power grid.Recently,there are several problems in AEC detection of power supply enterprises in China: (1)The AEC detection mainly relying on expert experience,establishing mechanism rule model identification,insufficient recognition ability and insufficient flexibility;(2)With the vigorous development of smart grids and the sharp increase in electricity consumption data,the complex structures and massive data streams have flooded a large amount of abnormal information and cannot be effectively detected;(3)Comparing to the normal case,the AEC data is relatively small,leading a typical imbalanced problem.And the traditional detection algorithm are not suitable for imbalance data learning.Therefore,it is urgent to develop intelligent and efficient abnormal electricity detection algorithms to ensure the smart power management.Based on the above,the research work of this paper is mainly divided into four parts:(1)Aiming at the complex and variable electricity consumption data,a novel feature extraction method that characterizes the electricity consumption data property based on the definitions of different AECs is firstly proposed for the electricity consumption data representation,such as electricity consumption fluctuations,which is referring to the idea of statistical learning method.(2)Aiming at the imbalanced distribution of abnormal electricity consumption,an Ensemble-ELM for AEC detection method based on the fusion of ELM and ensemble learning is proposed.Adopting ensemble learning can effectively solve the problem of imbalanced distribution of abnormal electricity data.With ELM as the base model,the abnormal electricity data has strong characterization ability,fast training speed and high recognition rate.Finally,a majority voting method is used to vote to further improve the accuracy of abnormal electricity data detection.(3)From the perspective of data balance,a K-means SMOTE + KELM algorithm for the AEC detection based on K-means SMOTE and kernel-based extreme learning machine(KELM)was proposed.The improved K-means SMOTE algorithm is used to synthesize a small number of abnormal electricity data,which can solve the problem of ignored the issues of within-class imbalance and generate noisy samples during the SMOTE sampling process.Finally,the ANN trained by the KELM algorithm is applied for feature learning and AEC detection,which can convert the original low-dimensional linear non-distributable electricity data into linear separable,and improve the recognition rate of abnormal electricity data.(4)From the perspective of data weighting,an AEC detection method based on a deep weighted ELM(DWELM)is proposed,which builds on an improved multiclass Ada Boost imbalanced learning algorithm(Ada Boost-ID)and an enhanced stacked multilayer deep representation network trained with the ELM(EH-Dr ELM).Among them,EH-Dr ELM can be used to improve the representation capabilities of features and useful information,and the improved Ada Boost-ID can further optimize the sample weights.The combination of the two structures can effectively solve the insufficient classification detection ability of general shallow layered classification algorithms for complex and imbalanced data.Finally,the simulation experiment results based on the electricity consumption data of State Grid Zhejiang Electric Power Corporation show that the proposed algorithms in this paper achieves a better performance than the traditional detection methods in AEC detection,which verifies the effectiveness of the proposed model.