Federated learning has emerged as a practical solution for residential load forecasting when privacy regulations prohibit centralized data collection. This research rigorously benchmarks whether grouping households into clusters—a common technique in federated learning—genuinely improves forecasting accuracy compared to simpler alternatives.
The study evaluates six different federated learning strategies across varying levels of data heterogeneity using two real-world smart meter datasets. A key innovation is the inclusion of personalization fairness: rather than comparing clustering only against a global model, researchers also benchmark household-specific fine-tuning as a baseline. This provides a more complete picture of clustering's true benefits.
Results reveal an important paradox. While clustering reduces forecasting errors by up to 26% when measured at the cluster-model stage, this advantage largely disappears once individual households fine-tune their models. After personalization, all methods converge within 2% of the global baseline, delivering nearly identical performance. Adaptive re-clustering—dynamically adjusting household groups during training—provides no accuracy improvement over static clustering approaches.
The paper's main contribution redirects expectations about clustering's value. Rather than improving accuracy, clustering primarily enhances communication efficiency, a critical metric in federated learning. Clustered frameworks converge using one-third the communication rounds required by standard global models, particularly important for portfolios with diverse distributed solar resources and heterogeneous consumption patterns.
These findings have practical implications for utilities and energy service providers deploying federated learning systems. Organizations may achieve accuracy targets without clustering complexity simply through allowing household-level personalization. However, when communication bandwidth is constrained—as in remote regions with limited connectivity—clustering offers meaningful efficiency gains justifying its implementation complexity.
The researchers released an open-source Ausgrid benchmark featuring a temporal out-of-time evaluation protocol, enabling standardized comparison of future federated load forecasting methods. This contribution supports reproducibility and accelerates research in privacy-preserving demand forecasting.
