Interpreting Atypical Conditions in Systems with Deep Conditional Autoencoders: The Case of Electrical Consumption

Abstract

In this paper, we propose a new method to iteratively and interactively characterize new feature conditions for signals of daily French electrical consumption from our historical database, relying on Conditional Variational Autoencoders. An autoencoder first learn a compressed similarity-based representation of the signals in a latent space, in which one can select and extract well-represented expert features. Then, we successfully condition the model over the set of extracted features, as opposed to simple target label previously, to learn conditionally independent new residual latent representations. Unknown, or previously unselected factors such as atypical conditions now appear well-represented to be detected and further interpreted by experts. By applying it, we recover the appropriate known expert features and eventually discover, through adapted representations, atypical known and unknown conditions such as holidays, fuzzy non working days and weather events, which were actually related to important events that influenced consumption.