Improving the accuracy of detecting non-technical electricity losses using sampling algorithms

On the basis of the obtained data of the Advanced Metering Infrastructure (AMI) on electricity consumption, machine learning methods were tested to determine the locations of non-technical electricity losses. The implementation of this methodology consists in predicting the probability of non-technical losses by time series of electricity consumers' load. Preprocessing of the data has been performed, which consists in filling in missing values, normalization, optimization of the features using the method of principal component analysis. An empirical study consisting of applying different sampling strategies (adjusting the training dataset) to solve the problem of dataset imbalance to improve the classifica-tion accuracy was carried out. The balanced random forest method was used with an accuracy on the test dataset using Roc-AUC metric of 0.8623. Based on the obtained classification results, the calculation of commercial loss reduction is made. The application of methodologies sampling methodology maintains the dynamics of a small increase in the ac-curacy of non-technical electricity losses detection. The small increase in the quality of models when using a balanced data set is explained by the fact that solving the problem of data set imbalance does not simplify the identification of sources of commercial electricity losses due to their similarity in some examples with normal consumption.