Federated electricity demand forecasting: A client selection mechanism based on TimeVAE latent representations

Artificial intelligence-based solutions are becoming commonplace in the reality of distribution networks. Low-level demand forecasting, in particular, faces significant challenges in large-scale operations and planning, especially regarding privacy concerns and computational performance. Federated Learning (FL), a decentralized training paradigm, emerges as a promising approach that meets these requirements. However, to ensure the success of this paradigm in highly complex environments, such as those involving distributed energy resources, FL solutions must consider a rational selection of clients. In such environments, full client participation may become computationally costly, making the approach infeasible, while random client selection may lead to incomplete pattern recognition. Other studies suggest clustering strategies relying on pre-existing data. However, in real-world applications, this data may not be available. Given these constraints, this paper presents a representativeness-based method for client selection. Using client-side TimeVAE models, the proposed method performs clustering by utilizing client latent spaces and identifies the best representative of the constructed clusters. Synthetic generation and clustering are evaluated by varying the dimensionality of the latent space and applying different distance metrics. Then, the proposed selection mechanism is evaluated in comparison to full and random client participation, adopted as baseline strategies. The outcomes demonstrate that the proposed method outperforms these baselines in terms of mean squared error, while preserving client privacy. Furthermore, the findings suggest the fidelity of synthetic data generation and the effectiveness of the proposed client selection, indicating its potential usage for power distribution applications with privacy concerns.