: ..,
..
:
: 113
: -
: 2025
: .., .. // . - 2025. - . 113. - .232-272.
: , , , ,
(.): point anomaly detection, typical daily energy consumption profile, machine learning, threshold algorithm, ensembles of algorithms
: . . 21 , . , . , , . , . , , contamination . . , ( contamination ), . , . , , .
(.): The paper describes studies on detecting point anomalies in energy consumption data using two different data sets as an example. Methods for constructing typical energy consumption patterns are considered and the authors' method for constructing a typical daily energy consumption profile is presented. To conduct numerical experiments, the authors selected 21 unsupervised machine learning methods suitable for detecting point anomalies. Based on the results of numerical experiments, the methods that most successfully coped with the task of detecting point anomalies were noted. Particular attention in the work was paid to methods that do not require additional parameters and modern, promising methods based on artificial neural networks. According to the test results, the best algorithms were statistical algorithms based on constructing histograms. One of the main problems addressed in the work is the problem of setting the contamination parameter for each considered algorithm. One of the solutions to this problem is the use of threshold algorithms. It is shown that if the original algorithm does not detect anomalies well enough (the contamination parameter is not configured), then the use of threshold algorithms can significantly improve the accuracy of anomaly detection. Threshold algorithms are noted, the use of which for the tasks of analyzing anomalies in energy consumption data, most often ensures an increase in accuracy. Threshold algorithms can be applied both to the results of individual anomaly detection algorithms and to the results of ensembles of algorithms obtained using various combination strategies.
PDF
: 72, : 38, : 29.