Appliance V-I Trajectory Detection for NILM based on Conditional Autoencoder.

Non-intrusive load monitoring (NILM) analyzes the terminal voltage and total current to offer comprehensive appliance-specific consumption data. Many state-of-the-art NILM models make a crucial assumption that recognized devices within the training set are responsible for triggering switching events. However, continuous addition of new devices diminishes the practical efficacy of existing methods for load monitoring. In this manuscript, we introduce a novel approach based on the Conditional Variational Auto-Encoder (CVAE) to address the challenge of classifying familiar appliances and identifying unfamiliar ones by leveraging V-I trajectory characteristics. In our proposed approach, during the training phase, we promote the alignment of capsule features belonging to the same familiar class with a predefined Gaussian distribution, where each class has its own distribution. To achieve this, we adopt the variational autoencoder framework and utilize a collection of Gaussian priors as an estimate for the following distribution. By employing this approach, we can regulate the compactness of features belonging to the same class around the mean of the corresponding Gaussian distributions. This enables us to control the classifier’s capability to identify samples from unfamiliar classes. Testing findings on the public dataset illustrate the efficiency of our approach.