Explainable AI and Ensemble Machine Learning Analysis of River Flow Dynamics: Influence of Key Climatic Variables (Temperature, Humidity, Precipitation)

Creative Commons License

Çaklr M., Ural G. N., ORAL M., ORAL O., YILMAZ M.

Journal of Hydrology and Hydromechanics, cilt.74, sa.2, ss.113-124, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 74 Sayı: 2
  • Basım Tarihi: 2026
  • Doi Numarası: 10.2478/johh-2026-0009
  • Dergi Adı: Journal of Hydrology and Hydromechanics
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, CAB Abstracts, Compendex, Environment Index, Geobase, Directory of Open Access Journals, Natural Science Collection (ProQuest), Earth, Atmospheric, & Aquatic Science Collection (ProQuest), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
  • Sayfa Sayıları: ss.113-124
  • Anahtar Kelimeler: Daily flowrate forecasting, Ensemble learning, Explainable AI, Hydrology, LIME, SHAP, Time-aware validation
  • Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
  • Akdeniz Üniversitesi Adresli: Evet

Özet

Accurate short-term river flowrate forecasting is essential for flood risk mitigation and sustainable water management. However, many machine learning (ML) applications in hydrology lack strict temporal validation and interpretability, limiting operational reliability. This study develops a reproducible and explainable framework for one-day-ahead daily flowrate forecasting in the Eşen River Basin (Türkiye) using hydro-meteorological data (2017-2022) from one flowrate station and four meteorological monitoring stations. The workflow integrates the Box-Cox transformation, lag-based feature engineering (up to 3 days, reflecting short-term hydrological memory), Boruta feature selection, and strictly time-aware rolling validation (2017-2021 training; 2022 independent test). Classical time-series models (ARIMA, TBATS), interpretable baselines (DT, LR), and advanced ML algorithms (RF, GBM, XGBoost, SVM, ANN) were benchmarked using RMSE, MAE, R2, NSE, and KGE. Ensemble tree-based models consistently outperformed classical and baseline approaches in magnitude-sensitive metrics. XGBoost achieved the highest predictive accuracy (R2 = 0.864; NSE = 0.864; RMSE = 6590 dm3 s-1). Although TBATS yielded the highest KGE (0.865), ensemble models better captured nonlinear dynamics and flowrate variability. SHAP and LIME analyses revealed that short-term flow lags dominate predictive structure, while precipitation and temperature exert regime-dependent influence. The complete workflow is openly deployed via a reproducible R-Shiny environment. The results demonstrate that explainable ensemble learning, combined with strict temporal validation, provides a reliable and transparent framework for operational hydrological forecasting.