Contrastive Learning with Class Collision Awareness for Periodic Forecasting in 6G Urban Digital Twins

Periodic time-series forecasting is central to 6G-enabled urban digital twins, where both cellular traffic management and environmental sensing demand accurate predictions over recurring diurnal and weekly regimes. Contrastive self-supervised learning has emerged as a promising approach for learning temporal representations, yet when applied to such periodic data, it suffers from class collision: temporally distant but semantically similar recurrent patterns are pushed apart as false negatives. We propose Clustering-Enhanced Contrastive Learning (CECL), which couples temporal contrastive learning with a clustering regularizer that maintains a Gaussian mixture structure over the latent space, preserving global periodic structure while retaining local discriminability. We evaluate CECL on five datasets across three tracks spanning two domains: cellular traffic forecasting (Milan CDR, 20 cells), multi-site hourly air-quality forecasting (Beijing Multi-Site, 12 stations; QUANT, 3 cities), and daily air-quality forecasting (Haikou and Taizhou). Across all tracks, CECL consistently outperforms supervised and contrastive baselines, reducing RMSE by 3–10% relative to the strongest contrastive competitor (CoST). These results demonstrate that clustering-guided contrastive regularization yields robust gains for periodic forecasting in both 6G network management and environmental sensing scenarios.