The conventional measures to identify power theft, electricity leakage, and electricity meter faults include periodic inspection, periodic electricity meter check, and user reporting. However, these measures depend on manual work and lack specific targets.
Currently, many power supply bureaus use the metering alert function and the power data query function to monitor users’ power usage online. The bureau staff identify power theft, electricity leakage, and electricity meter faults by collecting information such as abnormal power usage, electricity load exceptions, terminal alerts, primary site alerts, and line exceptions or losses. A model is created to analyze abnormal power usage based on the metric weights by collecting statistics on the current, voltage, and load at the metering point before and after an alert is triggered. This helps identify power theft, illegal power usage, and electricity meter faults.
The analysis model can collect information about abnormal power usage but fails to identify the users suspected of power theft or electricity leakage in a fast and accurate manner. This is a big challenge for the audit staff. The analysis model requires expert knowledge and experience to determine the weights of input metrics. This process is flawed and depends on subjective judgment, producing unsatisfactory results.
An automatic electricity metering system can collect statistics on electricity load, such as current in all phases, voltage, and power factor, as well as terminal alerts such as abnormal power usage. Alerts and electricity load data can reflect users’ power usage. The audit staff can identify users suspected of power theft and electricity leakage through an online audit system and onsite audit, and import findings to the system.
The imported data is analyzed to extract the key features of power theft and electric leakage and create a model used to automatically check and identify power thieves and households with electric leakage. This greatly reduces the workload of the audit staff and ensures normal and safe power usage.
You can select a dataset to view three metrics about a user’s power usage and the data that indicates whether the user steals power or encounters electric leakage. The three metrics are power usage decline level, line loss rate, and warning num. The “is theft” column lists the metric analysis result.
In the left-side navigation pane, choose Components > Statistical Analysis, and drag and drop Correlation Coefficient Matrix to the right section to view each feature related to the output power.
Right-click the completed component and select View Analytics Report to obtain the correlation analysis result. The correlation chart shows that the three metrics are not closely related to the result of “is theft”. That is, the features are not specific enough to determine whether a user is a power thief. Then, in the left-side navigation pane, choose Components > Statistical Analysis, and drag and drop Data View to the right section to analyze the distribution of data in the label column by feature. Select the feature columns as follows.
Then, select the label column.
Finally, right-click Run from Here, choose the completed component, and select View Analytics Report to view the distribution of data in the label column by feature.
After completing a simple exploratory analysis, you can select an appropriate algorithm model for data modeling. In the left-side navigation pane, choose Components > Data Preprocessing, and drag and drop Split to the right section to split data into the training set and test set.
Choose Components > Machine Learning > Binary Classification, and drag and drop Logistic Regression for Binary Classification to the right section to perform regression modeling on data. Select the feature columns (X) and label column (Y). In this experiment, the feature columns are power_usage_decline_level,line_loss_rate, and warning_num.
After modeling is complete, choose Components > Machine Learning and drag and drop Prediction to the right section to predict the effect of the model on the test dataset. For Feature Columns and Reversed Output Column, all options are selected by default. In the left-side navigation pane, choose Components > Machine Learning > Evaluation, and drag and drop Binary Classification Evaluation to the right section to view the model effect. The following figure shows the result of the experiment.
Right-click the Binary Classification Evaluation component to view the model effect. The AUC reaches the satisfying value 0.9827.
This completes the identification of power theft through Machine Learning Platform for AI (PAI). You can use Elastic Algorithm Service (EAS) to deploy the identification service so that it can be called online to identify power theft in power grids.
This experiment references the book “Python Practice of Data Analysis and Mining.” For copyright issues, contact the author of this topic. We respect every researcher in the academic field for their academic contribution and strive to better integrate technologies with the real life.