A Horizon-Adaptive Benchmarking Framework for Long-Term Reservoir Storage Forecasting Using Physics-Informed Transformers and Machine Learning
Water Resources Management, vol.40, no.1, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 40 Issue: 1
- Publication Date: 2026
- Doi Number: 10.1007/s11269-025-04399-w
- Journal Name: Water Resources Management
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, ABI/INFORM, Compendex, Environment Index, Geobase, INSPEC
- Keywords: Machine learning, Physics-informed transformer, Reservoir storage forecasting, Semi-arid basins, Time series forecasting
- Süleyman Demirel University Affiliated: Yes
Abstract
Accurate long-term reservoir storage forecasting is essential for sustainable water management in semi-arid regions. This study presents a horizon-adaptive benchmarking framework integrating physics-informed deep learning and classical time-series models for 12-month-ahead prediction. Five models—PIT-T, LSTM, SARIMAX, MLP, and an Ensemble—were trained on 12-month lags and evaluated across 12 horizons. PIT-T achieved the lowest 1-month RMSE (≈ 2.0 hm³) and best short-term memory preservation, MLP excelled at 2–6 months, LSTM dominated 7–10 months, and the Ensemble was most stable at 11–12 months, maintaining seasonal persistence. Robustness tests under reduced complexity and noise confirmed that PIT-T and Ensemble models retained accuracy (RMSE ≈ 5.5–5.8 hm³, NSE ≈ 0.22–0.29). These results show that no single model is optimal across all horizons, underscoring the need for horizon-specific model selection as a reliable, physically consistent approach for adaptive reservoir management in data-scarce basins.