Federated fuzzy k-means for privacy-preserving behavior analysis in smart grids

Abstract

Better understanding the behavior of various participants in smart grids, such as electricity consumers and generators, is important and beneficial for flexibility exploration and renewable energy accommodation. Clustering, as an effective data-driven approach to behavior analysis, has been widely applied for extracting the typical electricity consumption behavior of consumers and the bidding behavior of generators in smart grids. Traditionally, the clustering algorithms are implemented centrally with the assumption that data from all consumers or generators can be accessed. However, it may not be the case in the real world because the consumers and generators may not be able to or willing to share their own data due to privacy concerns or commercial competition. To address this issue, in this paper, we propose a federated fuzzy k-means method for privacy-preserving behavior analysis in smart grids. Specifically, two learning strategies, i.e., model averaging and gradient averaging, are designed for the implementation of the federated fuzzy k-means clustering. Both methods are investigated and comprehensively compared on both the electricity consumption behavior dataset and the generator bidding behavior dataset. Experimental results show that our proposed methods achieve similar performance to the traditional centralized fuzzy clustering method on independent and identically distributed (i.i.d.) data, as well as protecting the privacy of different participants in smart grids. As for non-i.i.d., the performance of the model averaging-based method worsen; in contrast, the gradient averaging-based method is more robust to this situation.