Research And Application Of Power Theft Detection Based On Data Mining

Power theft detection refers to the process of analyzing the user load data summoned by the power consumption information acquisition system through data mining,and finally outputting the users suspected of stealing electricity.First of all,the significance of power theft detection is that the on-line data analysis replaces the on-site visit and census,which saves the manpower and material resources of the power supply company.In addition,advanced technical analysis means greatly improve the efficiency,accuracy and coverage of theft detection,making power theft investigation more real-time,accurate and comprehensive.In this paper,two kinds of power theft detection models are constructed around the outlier algorithm suitable for large-scale user clusters and the neural network algorithm suitable for small-scale important users.Applied to the anti-theft and violation survey action of the power supply company and the inspection action of the important users in the direct supply area,good practical results have been achieved,with high precision and recall,and the cumulative electricity charge loss of more than 1.5 million yuan has been recovered for the power supply company.In the work of power theft detection based on outlier algorithm,the regular monthly load data is used as the benchmark to construct the high-dimensional feature index of power theft detection,and then the intensive index dimension of principal component analysis is used to make the user data visible in low-dimensional space.the influence of redundant information and noise information is eliminated.As for the outlier algorithm itself,the detection object is pruned by the introduction of index space and upper bound value,which reduces the time complexity of the outlier algorithm and makes it more adaptive to the calculation of large-scale data clusters.In the detection of electricity theft based on neural network algorithm,in addition to the basic index electricity volatility and its related follow-up indicators,season,energy consumption and other categories of indicators are also covered,in order to more comprehensively reflect the actual power consumption of users.As for the neural network algorithm itself,the structure,learning rules,initial weights and thresholds of the network are optimized by genetic algorithm,so that the network can automatically select,crossover and mutate according to the sample information,reduce the possibility of generating local solutions,and effectively improve the convergence speed and detection accuracy of the network.