Multi-Model Reservoir Occupancy Projection Framework for Istanbul: A Traceable Forecasting and Climatology-Anchored Simulation Workflow Integrating FAO-56 Evapotranspiration, Machine Learning, and Water-Balance Hydrology
Makaleler > Tam Makale · 2026
Yazarlar: Mehmet Yasin KAYA
Zenodo (CERN European Organization for Nuclear Research), Uluslararası, 2026
Özet
Reservoir storage is a central component of municipal water security in Istanbul, so shifts in hydroclimatic conditions and water demand can quickly propagate through the system. This study introduces a traceable monthly workflow that links institutional reservoir records from ISKI/IBB with curated Kandilli climate inputs for both forecasting and projection analysis. The framework combines three layers: a provenance-aware harmonization pipeline, a 20-model benchmark covering linear, tree-based, probabilistic, time-series, and physics-based comparators, and a physics-informed projection module built on FAO-56 water-balance calculations. To limit temporal leakage, model comparison relies on purged walk-forward cross-validation with a 6-month gap between training and validation windows; model behavior is also illustrated through a fixed-interval holdout diagnostic with 5-year and 10-year training histories. Cross-validation results are tightly clustered, with the best models falling between 39.33 and 40.52 percentage points root mean squared error (RMSE). The holdout diagnostic yields lower short-horizon errors (best 5-year RMSE = 4.86; best 10-year RMSE = 5.83), butthose values are interpreted only within that separate protocol because the target definition differs from the cross-validation setting. Missing driver months in 2025 are completed with month-of-year climatological averages, and the 2026–2040 baseline extends the same drivers through deterministic seasonal climatology rather than a multi-GCM scenario ensemble. Under this fixed forcing design, the system-wide 2040 ensemble median occupancy reaches 16.3%, with a P20–P80 inter-model spread of 10.1–27.8%. The Elmalı-focused module ends lower in both central tendency and endpoint range than the aggregated system forecast.