Federated Online Learning for adaptive load forecasting across decentralized nodes

Load forecasting is vital for a country’s power industry and economy, playing a crucial role in infrastructure planning, grid operation, and resource allocation. In recent years, Machine Learning (ML) techniques have been dominating load forecasting; however, traditional ML requires transferring data from different sensors, such as smart meters, to a central location for model training, and then the static models are used for forecasting. The main drawbacks are twofold: (1) data sharing and transfer results in privacy and security risks, and (2) static models miss opportunities to learn from new data. Federated learning has been proposed to address the first drawback while online learning tackles the second one. However, integrating the two remains challenging as it requires reconciling the distributed yet static paradigm of FL with the centralized yet dynamic sequential updating characteristics of online learning. Consequently, this paper proposes Federated Online Learning (FOL) which reconciles federated and online learning to provide adaptive distributed load forecasting. The local nodes employ a modified online learning technique based on a Long Short-Term Memory (LSTM) to learn from streaming data, while the federated learning process, including aggregation, is designed to accommodate sequential learning. Moreover, FOL includes communication control parameters to manage the exchange of information, synchronization, and network traffic. Experiments demonstrate that the proposed FOL outperforms traditional offline models and achieves similar results to online models while providing the benefits of adaptive distributed learning without local data sharing. • Traditional machine learning and online learning require sharing data with the server. • Federate learning (FL) trains without data sharing but produces static morels. • Our Federated Online Learning (FOL) combines the benefits of online learning and FL. • FOL learns from streaming data in a distributed manner. • FOL outperforms offline models and matches online models without sharing local data.