A low-level set of stationary features dedicated to non-intrusive load monitoring

We study a NILM (Non-intrusive load monitoring) problem by considering the classification in two steps: a first step centered on the short duration signal of an electrical device (around one second) and a second step centered on the sequence of operating modes of this electrical device (separated by several seconds). The second step is mainly based on an RNN (recursive neural network) and will be described in future work. In this article, we study the first step which consists of a simple classification of certain electrical devices in stationary mode. We show that two parameters and some harmonics constitute a characteristic set of features and we look for a minimal set of features that preserve good results. We compare (both theoretically and practically) the meaning of odd and even harmonics. We illustrate our results using the PLAID database and some of our own electrical devices, in order to validate our real-time embedded system.