Guard Framework Improves Scientific Time Series Forecasting with Multi-Teacher Distillation

Deploying large Time-Series Foundation Models (TSFMs) in scientific fields faces two main hurdles: their high computational cost for edge devices and their poor zero-shot performance when applied to specific scientific domains due to data distribution shifts. This research introduces Guard, a novel framework designed to extract valuable structural knowledge from these powerful but often misaligned foundation models to train smaller, specialized forecasters. Guard re-imagines multi-teacher distillation as an instance-wise decision process. It incorporates two adaptive mechanisms: a Contextual Router that dynamically selects the most relevant teacher model based on local input characteristics, leveraging the complementary strengths of diverse FMs. Additionally, an Uncertainty-Gated Temperature mechanism acts as a "circuit-breaker," reducing distillation intensity when a teacher's confidence diverges significantly from the target domain's reality. Evaluated across critical climate domains like meteorology and energy grids, Guard significantly reduces prediction errors compared to fixed-weight multi-teacher distillation. It demonstrates that even misaligned foundation models can serve as crucial correctives, improving performance on challenging instances and enabling high-precision forecasting suitable for resource-constrained edge deployments.