Researchers have developed a personalized federated sparse adaptation framework for time-series foundation models (TSFMs), aiming to improve energy forecasting by addressing the non-IID nature of private, distributed meter data. This new approach utilizes a heterogeneous temporal mixture-of-experts (MoE) adapter, where a sequence-level router selects a subset of experts tailored to specific context windows. Experiments across 50 buildings and three TSFM backbones demonstrated that this personalized strategy consistently outperforms global federated learning and local adaptation methods, highlighting the importance of client-aware and backbone-aware adaptation. AI
IMPACT This research could lead to more accurate and privacy-preserving energy forecasting by improving how distributed, non-IID time-series data is utilized.
RANK_REASON The cluster contains an academic paper detailing a new adaptation framework for time-series foundation models.
Read on Hugging Face Daily Papers →
- Federated Adaptation
- Global Flatline
- Hugging Face
- Innu-aimun
- Local MoE
- mixture of experts
- Time Series Foundation Models
- TSFMs
- arXiv
- Pandarasamy Arjunan
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
