Use of Binary Classification in Non-Invasive Load Monitoring

The increasing energy intensity of the economy has led us to look for ways to reduce this negative trend. One method is non-intrusive load monitoring (NILM). This paper presents the use of artificial intelligence methods for the selection of information features and for the identification of operating electrical devices. A set of potential identification features was obtained from high-frequency measurements covering 12 types of electrical consumers and consisted of 218 features. From these, an identification vector was selected via the mRMR (minimum redundancy maximum relevance) method, which searches for features that are maximally correlated with the class and are as little correlated with each other as possible. Identification was realized by building a hybrid classifier using binary classifiers built from artificial neural networks and decision trees. The Accuracy, Precision, Recall, and F1 metrics were used to assess the quality of identification. The obtained values of the identification quality indicators confirm that it is possible to replace multiclass classification in NILM with binary classification without losing its quality. The use of binary classifiers allows for the identification of new devices without the need to change the classifier configuration.