Work

GLINT: Multi-seasonal Residual Gating in a Linear-Transformer for Cross-dataset Multi-horizon Water Demand Forecasting

Year: 2026
Type: article
Abstract: Multi-horizon water demand forecasting must preserve recurrent demand patterns while adapting to heterogeneous residual dynamics.This study proposes GLINT, a multi-seasonal residual gating framework combining explicit daily and weekly seasonal experts with linear and multi-scale Transformer residual experts.A stepwise gate conditioned on residual statistics, Transformer context, and known-future calendar features fuses the four forecasts.GLINT is evaluated on 10 BWDF and 11 SOWEKI operational series at 24-, 72-, 168-, and 336-hour horizons using chronological splits, causal missing-value treatment, and three random seeds.Across the evaluated baselines, GLINT achieves the lowest MASE in all eight dataset-horizon settings, improving the strongest baseline by 7.01 to 12.93 percent.Under Holm correction within each dataset-horizon family, 30 of 40 comparisons remain significant.Ablation and sensitivity analyses show that seasonal residualization, the linear expert, and adaptive gating are consistently useful, while Transformer correction provides dataset-dependent gains.
Language: English
Open Access status: hybrid
OpenAlex ID: W7234879131
FWCI:
Cites: 38
Cited by:
Related to: 10