Enhancing Short-Term Energy Forecasting in Distributed Systems through Federated Learning

Abstract

Energy consumption forecasting plays a crucial role in smart buildings, enabling energy monitoring, planning and balance between supply and demand. Especially nowadays it is a useful tool for encouraging consumers to change their daily habits in a manner that positively impacts the environment. Public educational institutions are one of the most appropriate study cases because they are research centers that develop and create novel concepts and serve as examples for energy−efficient buildings. An additional challenge is the development of a global model without centrally collecting and revealing raw data. This paper proposes a generalized solution for effective short−term energy consumption forecasting in buildings at the University of Cyril and Methodius in North Macedonia. The proposed model is trained on 25 different university department buildings’ data from January 1, 2021 to November 16, 2023 with resolution of one hour without compromising data privacy. Dataset contains energy consumption and extracted calendar variables as hour, weekday and month. The proposed approach utilizes federated learning technique to develop a generalized deep learning model for multi- step forecast that operates efficiently and enhances interoperability among different energy consumers. Additionally, a comparative analysis between recurrent neural network in centralized and decentralized learning approach is presented. The model performances are evaluated using the following metrics: RMSE and MAE. The obtained results show that federated learning achieves great performance.