Client clustering versus personalization in federated residential load forecasting: A personalization-fair benchmark of accuracy and communication overhead

Residential load forecasting underpins demand response and local energy trading, yet privacy regulations such as the EU GDPR prevent service providers from collecting fine-grained household consumption data. Federated learning (FL) enables collaborative training without raw data exchange, and client clustering widely mitigates accuracy degradation caused by non-IID residential load distributions. However, existing clustering benchmarks only compare grouped schemes against a single global FL model, ignoring household-specific fine-tuning as a strong personalized baseline. This work constructs a personalization-fair benchmark to quantify the true value of client clustering for residential load forecasting. Six FL strategies are evaluated across three data heterogeneity regimes built on two real-world smart meter datasets. Results show clustering reduces forecasting errors by up to 26% at the cluster-model stage, yet after household fine-tuning all methods land within 2% of the global model and deliver nearly identical performance; adaptive re-clustering brings no accuracy improvement over static grouping. The primary merit of clustering lies in communication efficiency, as clustered frameworks converge within one-third of the communication rounds required by the standard global model under highly heterogeneous PV prosumer portfolios. Finally, this paper releases an open Ausgrid benchmark with a temporal out-of-time evaluation protocol for follow-up federated load forecasting research.