Personalized Federated Sparse Adaptation of Time-Series Foundation Models

Research Question and Approach

This paper addresses the challenge of adapting pretrained Time-Series Foundation Models (TSFMs) for building energy forecasting in a federated learning (FL) environment. Because building energy data are private and highly non-IID (independent and identically distributed), standard global model aggregation often fails to capture building-specific temporal dynamics. The authors propose a personalized federated sparse adaptation framework that inserts a lightweight Mixture-of-Experts (MoE) adapter after the frozen backbone of a TSFM. This adapter uses sequence-level routing to assign input windows to a subset of experts, each designed to capture specific temporal inductive biases like periodicity, trend-residual structure, and multi-resolution behavior.

The study evaluates three TSFM backbones (MOMENT-1-large, Chronos-2, and Moirai-1.1-R-small) across 50 buildings. The results demonstrate that personalized MoE adaptation significantly reduces NRMSE compared to global FL-MoE, with improvements ranging from 7.1% to 12.5% depending on the backbone. The authors find that the effectiveness of the adaptation is backbone-dependent: for instance, MOMENT exhibits client-level expert specialization, while Chronos-2 shows high expert concentration, and Moirai displays more uniform routing. This indicates that federated personalization must be designed with the specific architecture of the underlying foundation model in mind.

As TSFMs become standard for time-series forecasting, their deployment in privacy-sensitive domains like energy management requires robust federated adaptation strategies. This work provides a blueprint for balancing global knowledge transfer with local personalization. By using sparse MoE adapters, the framework not only improves predictive accuracy but also reduces communication overhead, making it a practical solution for resource-constrained edge devices in smart buildings.