• Couples temporal augmentations with graph-aware positives to learn robust embeddings for nonstationary signals.
• Improves downstream forecasting accuracy under dynamic topology changes.