Time series prediction & generation from disentangled latent factors: new opportunities for smart cities

The acceleration of urbanisation has brought many new challenges to cities around the world. Application range is wide, from air pollution to public transportation modelling. The availability of data pertaining to these issues has been growing fast in the last years, offering many opportunities to tackle those applications with machine learning algorithms. We propose an elegant and general architecture that is able to provide state of the art forecasting in several different domains. Our idea is the following: for many time-series, a number of factors, that often relate to the context they were created in, can influence the observed values, such as day or location. In this paper, we present a machine learning model that learns to represent and disentangle such factors. Our contribution is to provide an approach that works at different scales: on a short term basis (30 minutes to few hours) our deep neural network architecture delivers competitive forecasting in a classical setting; at the day/week/month level, we show that we can generate relevant time series associated with unknown contexts. To the best of our knowledge, this ambitious application has not been investigated until now.