Personalized Federated Sparse Adaptation of Time-Series Foundation Models
The paper presents a personalized federated adaptation framework for building energy forecasting. It places a heterogeneous temporal mixture-of-experts adapter after a pretrained time-series foundation model representation. A sequence-level router selects a top-k subset of experts for each 168-hour context, with experts targeting periodicity, long-range interactions, local variation, trend-residual structure, and multi-resolution behavior. Experiments across 50 buildings and three TSFM backbones report that personalized federated variants consistently outperform global federated MoE and local training, while the strongest sparse-adaptation strategy depends on the backbone and evaluation metric.
Why it's worth reading
Private, heterogeneous building-meter data are a concrete barrier to deploying time-series foundation models. This work evaluates shared, local, personalized, and sparse expert adaptation together, making its trade-offs immediately relevant to federated forecasting systems.