REDNet: Two-stage knowledge distillation for robust edge time series forecasting

Edge devices are inherently constrained in computation, memory, and energy, creating a conflict for time series forecasting: large-scale models offer robustness but are infeasible to deploy, while existing lightweight models are efficient but often struggle with noisy or corrupted data. To address this limitation, we propose the Robust Edge Distillation Network (REDNet) , a framework that transfers the deep temporal understanding and robust generalization of a large pre-trained foundation model to a compact, edge-efficient student model. REDNet augments target-dataset supervision with training signals from a pre-trained teacher, while retaining only the compact student for deployment. The framework combines an MLP–BiGRU student with a two-stage distillation procedure that first aligns intermediate representations and subsequently performs end-to-end task adaptation. Extensive experiments on eight multivariate benchmark datasets demonstrate that REDNet consistently outperforms all lightweight state-of-the-art baselines in forecasting accuracy. Critically, our distilled model achieves accuracy and robustness-to-corruption levels that are comparable to the large pre-trained teacher model, all while using only 174k parameters and 3.8% of the teacher’s computational cost (MACs).