Bayesian-optimized machine learning for discharge prediction in the Alata River Basin (Türkiye): meteorological station representativeness
Hydroclimatic volatility across the Mediterranean Basin increases the need for reliable river-discharge prediction, particularly in data-scarce coastal catchments where random validation and untested station representativeness may overstate skill. This study develops a Bayesian-optimization workflow for the Alata River Basin, Türkiye, using meteorological predictors and one-day lags of precipitation and air temperature. Gaussian Process Regression, Support Vector Regression, and Multi-Layer Perceptron models, together with BO-tuned variants, were evaluated using a chronological 80/20 split and time-ordered cross-validation. Station-specific datasets showed that data source and record completeness changed predictive accuracy and model ranking. At MERS, BO-MLP achieved the highest test R2 (0.889), while BO-SVR yielded the lowest MAE. At ERDM, baseline SVR performed best overall (R2 = 0.869), whereas BO-MLP and BO-GPR were the strongest optimized configurations. A three-scenario Monte Carlo experiment with 30 replications per scenario showed robust gains for SVR and MLP, while GPR improvements were smaller and condition-dependent across controlled data conditions.
Authors
- Ümit Yıldırım (ORCID: https://orcid.org/0000-0002-7631-7245)
- Didem Güleryüz (ORCID: https://orcid.org/0000-0003-4198-9997)
- Cüneyt Güler (ORCID: https://orcid.org/0000-0001-8821-6532)
- Onur GÜVEN (ORCID: https://orcid.org/0000-0001-5608-7633)
Institutions
- Bayburt University (TR)
- Mersin Üniversitesi (TR)
Publication Details
- Journal
- Hydrological Sciences Journal
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1080/02626667.2026.2756497
- Primary Topic
- Hydrological Forecasting Using AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00