Research On Electricity Theft Detection Based On Sequence Similarity Discrimination And Ensemble Learnin

The stable operation of the power industry guarantees people’s daily safe use of electricity.The progress and development of society also promotes the rapid development of the power industry.Through the analysis and research of the massive load data of users,abnormal power consumption behaviors can be found in time to eliminate potential safety hazards.Preventing the occurrence of electricity theft is of great significance for ensuring the safe and stable use of electricity for the people,making the entire power system run smoothly and well,reducing economic losses,and ensuring life safety.This paper starts from the real load data that reflects the user’s electricity consumption behavior,and focuses on the detection of electricity theft under the analysis of the user’s electricity consumption behavior.It is convenient for the model to learn better detection.Through data preprocessing simulation enhancement and cluster calculation similarity,the electricity stealing sequence in the load data is screened and identified,and the corresponding features are extracted while the electricity stealing sequence is initially screened,and the electricity stealing discriminant features are analyzed and calculated.After screening,the idea of integrated learning is adopted.Detect the electricity stealing sample set.The main research content and specific work of this paper are as follows:(1)Stealing sample simulation and enhancement.Aiming at the problem that electricity stealing samples are difficult to collect and the number of samples is small in real scenarios,the electricity stealing sample set is simulated by analyzing the data characteristics of different electricity stealing methods,and according to the extremely unbalanced positive and negative samples of normal and electricity stealing in real scenarios The problem is to use Generative Adversarial Network(GAN)to enhance the power-stealing samples to balance the sample data set.Compared with other data sampling methods,the effectiveness of the method used is verified,and the powerstealing sample set used in the subsequent similarity screening and integrated learning lay the foundation for the construction.(2)Load sequence similarity discrimination.Aiming at the improvement of load sequence similarity discrimination accuracy and electricity stealing feature extraction,based on the simulation and enhancement of electricity stealing samples,information entropy is used to reconstruct the load data,highlighting the characteristics of the time when hidden dangers occur,and then the improved piecewise linear representation(PLR)algorithm is used to analyze the load data.Dimensionality reduction processing is performed on the sequence to improve the comparison efficiency while highlighting the characteristics of load peak changes.The sequence after dimensionality reduction is clustered into 4 categories using Canopy to determine the initial clustering center and cluster number of Fuzzy C-means(FCM)clustering,and extract each clustering center,that is,the load characteristic curve and the load sequence of the user to be tested for Dynamic Time Warping(DTW)distance calculation.And transform the distance into similarity to screen and judge the electricity-stealing sequence.After comparing different thresholds,the similarity threshold of 0.7 is used to effectively screen the suspected electricity-stealing sequence.Information entropy and effectiveness of improved PLR dimensionality reduction methods.(3)Ensemble learning for electricity tampering detection.Aiming at the problem of improving the performance of the electric stealing detection classifier,on the basis of the load stealing sequence screening,13 relevant features reflecting the characteristics of the load sequence change are selected and calculated,and RF is used to filter the feature set,which is consistent with the original 48-point sampling features.Constitute a feature set,use the method of ensemble learning to detect electricity theft on the load data set,and selectively integrate the base learners in the integrated model,and use Bagging heterogeneous ensemble learning with improved integrated voting strategy to conduct electricity theft sample data set Detection and recognition,compared with other single class learners and ensemble class learners,verified the effectiveness of the proposed model.