Federated Learning and its application in Smart Grid

Abstract

Federated Learning stands as a pivotal innovation within the landscape of machine learning, offering a transformative approach to model development by addressing the intricate balance between data utility and user privacy. In the context of contemporary data-driven research and applications, the traditional paradigm of centralising vast datasets for model training poses significant challenges, such as privacy concerns and regulatory compliance. This thesis delves into the algorithms and related applications in the smart grid. The work presented in this thesis mainly focuses on Horizontal Federated Learning, where participating clients have similar data structures, from three perspectives: i) presenting federated learning algorithms to enhance local consistency under data heterogeneity problem, ii) developing a federated learning framework for individual load forecasting for the distribution network, and iii) describing the personalised federated learning framework in household profile identification for different retailers. Results are presented to demonstrate the effectiveness of the proposed algorithm and framework. In addition to algorithm design and experimental results, this thesis also presents the current state of research on federated learning and overviews its development in the smart grid area. Based on the work presented, a multi-step inertial federated learning algorithm is first proposed for the data heterogeneity problem by utilising historical information. This approach effectively improves local consistency, thereby mitigating the impact of heterogeneous datasets. Then, individual probabilistic load forecasting based on federated learning for electricity consumers to forecast their domestic load is set up to protect users’ privacy, This approach has taken into account different time resolutions of smart meters, and quantity skew of different consumers. The active consumers are selected in each round, which can be against communication failure. Finally, the personalised federated learning household profile identification framework for retailers is put forward to handle the problem of personalised federated learning in household profile identification. A hyper network-based personalised federated learning approach is introduced for the coordination server. This method generates individual models tailored to the specific needs of each participating retailer.

Date of Award	14 Jun 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Zhirun Hu (Co Supervisor) & Zhengtao Ding (Main Supervisor)