Research On Power Consumption Anomaly Detection And Forecast Based On Machine Learning

Abnormal power consumption behaviors not only could affect the revenue of power supply companies,but also the electric larceny caused by the replacement of metering devices can damage the power infrastructure,which further threatens the safe and stable power supply of the whole power grid.Abnormal detection and consumption prediction have important practical significance for energy resource management and control.On the one hand,detecting abnormal power consumption of the power grid is very important for the stable operation of the entire power grid,and after detecting abnormal users,determining the actual power consumption of the users is of guiding significance for the power supply company to handle related matters.On the other hand,consumption forecasting based on historical data is essential to improve the energy service capabilities of users.In order to detect abnormal users,a short-term fraud detection method based on the Support vector machine(SF-SVM)is proposed in the thesis.First,the system automatically collects the recent limited-length power data of the regional power grid and users.When the system detects that the difference between the power consumption provided by the regional power grid and the power consumed reported by the smart meter exceeds a certain level,the system will change to the electric larceny suspicious status,and trigger fraud detection process.After that,system will compare the user's electricity consumption data before and after the detection of electric larceny,then extract the feature values from the relevant data,thereby identifying each user one by one and finding out suspicious electric larceny users.The comparison experiment results show that SF-SVM has a higher detection rate and a greater use value when the false alarm rate is low.In order to predict the power consumption of customers,a customer power consumption prediction method combining by the complete ensemble empirical mode decomposition with adaptive noise algorithm and long short-term memory algorithm(CEEMDAN-LSTM)is proposed.First,CEEMDAN is introduced to decompose the trend component,periodic component and random component of electric load time series data for solving the problem that it is difficult to mine deeper time-series features of users' power consumption data by applying the existing prediction models directly for nonstationary and non-linear.Then,based on the CEEMDAN,LSTM forecasting sub model is constructed,which can predict the total power consumption by the superposition.Finally,multiple sets of comparative experiments prove that the prediction error of CEEMDAN-LSTM is smaller.