• edge computing
• energy consumption forecasting
• Federated Learning
• LSTM
• residential buildings

With the advent of smart meters and advanced metering infrastructures, large size of energy consumption data has become available for residential buildings. This data can be used to train deep learning models for applications that want to monitor electrical load and demand response. These applications, however, require that the data have to be sent to a central server creating privacy issues. On the other hand, deep learning requires a large amount of data to be trained properly. In this thesis, a decentralized architecture is proposed to overcome the problem of data privacy. Federated Learning and the computational capacity of Edge Computing Nodes are used to train LSTM models. Different scenarios have been analyzed, in one of which the Porta a Lucca district of Pisa is taken as reference, changing the configuration of the Edge Computing Nodes and the number of Smart Meters connected to each of them. It has been shown that similar performance forecasting can be achieved in terms of RMSE. Furthermore, the results have been analysed in order to provide the disadvantages and advantages of using the federated approach. The simulation was done using Flower (Federated Learning framework) on data containing information on up to 328 houses. This thesis has been carried out within AUTENS (Sustainable Energy Autarky), a research project funded by the University of Pisa under the program PRA2020.