Federated learning framework for prediction of net energy demand in transactive energy communities

The implementation of transactive energy systems in communities requires new control mechanisms for enabling end-use energy trading. To optimize the operation of these communities, the availability of accurate predictions for the net energy demand is fundamental. However, to ensure effective management of flexible resources, the local generation and demand must be foretasted separately instead of just forecasting the net-energy demand. Additionally, to improve the forecast systems, more detailed data from the buildings are needed, but most information (such as patterns of occupancy) can be private. This paper proposes a novel federated learning (FL) framework for predicting building temporal net energy demand in transaction energy communities. The proposed approach is based on an FL architecture and has two independent forecast systems (generation and demand systems), ensuring collaborative learning among the buildings without sharing private data. The developed framework allows the integration of third-party data providers and facilitates coordination by a central server. The main goal of the framework is to support the management systems of transactive energy communities by computing the forecast of demand, generation, and net-energy demand. Additionally, such a framework has the novelty of introducing as an auxiliary system of Federated Transfer Learning, which will guarantee a more capable forecast system for new communities. The developed structure was tested using two communities, one with 100 buildings and the second with 25. The results showcase high accuracy and adaptability to different variables and scenarios, for instance, seasonal variations.