Expert Complementarity and Residual Correction for Coastal Sea-Level Forecasting
Abstract Coastal water levels vary with tides, weather, and ocean conditions. We study forecasts of the non-tidal residual over the next 24 h. Two Graph WaveNet experts use the same historical inputs but learn different targets. One predicts the residual alone; the other also predicts current and wave states. HS-DT combines their forecasts with a fixed rule. C4 then adds a bounded correction and a predictive uncertainty estimate while keeping the experts frozen. We compare these methods with deep and linear baselines at seven northeastern U.S. stations. Further tests examine later periods and a separate ten-station network. HS-DT improves sequence forecasts in several temporal evaluations. C4 improves historical sequence accuracy and probability scores, with most of the point-forecast gain coming from residual correction. Its component effects change in the later backtest. VARX remains a strong baseline, and the tested physical priors give small, mixed gains. These results support expert combination and bounded correction, while showing that their benefits depend on the evaluation setting.
Authors
- Yu Wang (ORCID: https://orcid.org/0000-0001-6798-2721)
- Guanchao Tong (ORCID: https://orcid.org/0000-0002-3503-2745)
- Yiling Yang
- Enze Cai
Institutions
- Northwestern University (US)
- Wenzhou-Kean University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-11531941/v1
- Primary Topic
- Hydrological Forecasting Using AI
- Type
- preprint