Impact of Federated Learning On Smart Buildings

Advances in Internet of Things open up many new possibilities in the design of smart-buildings. Thanks to the wide deployment and the growing computing power of connected devices, the researchers can now consider in-situ with distributed learning processes where on connected devices, local learners collect data to improve the building management system knowledge. In-situ and distributed learning is a convincing direction for smart-buildings since they can better take into account users contexts, privacy concerns or the minimization of energy consumption. However, this direction introduces new challenges such as that of formulating distributed learning processes in which data are collected locally, aggregated to generate models which themselves serve to define a global intelligence. This paper considers the formulation of such a distributed learning process, Federated personalization: a distributed learning strategy where individual learners collaborate in a knowledge sharing round to benefit from shared generalization, yet retain local specificity. This research work proposes a federation framework to orchestrate a group of autonomous learners naturally distributed across smart buildings. The results show that leveraging inter learner knowledge transfer makes it possible to achieve personalized lower generalization loss for regression problems like forecasting time series data and image classification. The prototype is currently deployed at Qarnot, a French computing heater company to provide secure and sustainable autonomic ambient intelligence via edge devices.