Houmma et al. (2026) Seasonal forecasting of dam water resources using optimized hybrid models under unprecedented drought conditions

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-01-06
• Authors: Ismaguil Hanadé Houmma, Abdessamad Hadri, Abdelghani Boudhar, El Mahdi El Khalki, Ismail KARAOUI, Sabir Oussaoui, Mohamed Samih, Christophe Kinnard
• DOI: 10.1016/j.ejrh.2025.103091

Research Groups

• Department of Environmental Sciences, University of Qu´ebec at Trois-Rivieres, Trois-Rivières, QC, Canada
• Research Centre for Watershed–Aquatic Ecosystem Interactions (RIVE), University of Qu´ebec at Trois-Rivières, Trois-Rivières, QC, Canada
• International Water Research Institute, Mohammed VI Polytechnic University (UM6P), Benguerir, Morocco
• L3G Laboratory, Faculty of Sciences and Techniques, Cadi Ayyad University, Marrakech, Morocco
• Data4Earth Laboratory, Sultan Moulay Slimane University, Beni Mellal, Morocco

Short Summary

This study developed optimized explainable artificial intelligence (XAI) models for monthly forecasts of water resource variations at the Al Massira dam in Morocco. It found that Bayesian probabilistic Long Short-Term Memory (ProbLSTM) and Generalized Additive Models (GAM) consistently outperform Light Gradient Boosting Machine (LightGBM) for seasonal forecasting, especially under unprecedented drought conditions.

Objective

• To develop XAI model pipelines utilizing Bayesian optimization via Optuna and objective integration of potential predictors.
• To compare the predictive scores of different models for forecasting the monthly water volume of the Al Massira dam up to six months ahead under two distinct scenarios: near-normal hydroclimate conditions and unprecedented changes due to severe prolonged drought.
• To determine the partial and overall contributions of predictors to the predictability of the monthly water volume of the Al Massira dam.

Study Configuration

• Spatial Scale: Oum Er Rbia watershed, Morocco, focusing on the Al Massira Dam (approximately 35,000 km²). Reanalysis data were extracted for the upstream part of the dam (altitude > 1500 m).
• Temporal Scale: Monthly forecasts for 1 to 6 months ahead. Historical data for dam water volume from January 1983 to July 2023. Predictor data are monthly, with lags up to 3 months. Forecasts were evaluated for the last 6 months of 2020 (near-normal conditions) and the last 6 months of 2023 (unprecedented drought conditions).

Methodology and Data

• Models used:
  • Light Gradient Boosting Machine (LightGBM)
  • Generalized Additive Models (GAM)
  • Bayesian Probabilistic Long Short-Term Memory (ProbLSTM) networks
  • Optimization: Bayesian optimization via Optuna for hyperparameter tuning.
  • Feature Selection: Least Absolute Shrinkage and Selection Operator (LASSO) regression for LightGBM and GAM; Minimum Redundancy Maximum Relevance (mRMR) method (using Joint Mutual Information criterion) for ProbLSTM.
  • Forecasting Strategy: Direct-recursive (DirRec) approach for 6-month-ahead forecasts.
  • Cross-validation: Time Series Cross Validation (TSCV) with 4 splits.
  • Model Interpretation: Shapley Additive Explanations (SHAP) for LightGBM; Integrated Gradients (IG) for ProbLSTM.
• Data sources:
  • Target Variable: Monthly total water volume of the Al Massira dam (January 1983 - July 2023) from the Oum Er-Rbia Hydraulic Basin Agency.
  • Large-scale climate teleconnection indices: Monthly time series (Southern Oscillation Index (SOI), Pacific Decadal Oscillation (PDO), North Atlantic Oscillation (NOA), Nino1+2, Nino3, Nino4, Pacific/North American index (PNA), Oceanic Niño Index (ONI)) from the Climate Prediction Center.
  • Hydrometeorological variables:
    • ERA5-Land monthly aggregated climate reanalysis data (11.1 km spatial resolution) from ECMWF: Volumetric soil water layer 4 (VSW4), Runoff max (Rmx), Subsurface runoff sum (SSRS), Total precipitation sum (TPS), Maximum totalevaporation (TEmx), Maximum snow melt (Smeltmx).
    • NASA Prediction of Worldwide Energy Resources (POWER) platform: Root zone soil moisture (RZSW), Surface soil wetness (SSW), Profile soil moisture (PSM), Relative humidity at 2 m (RH2M).
    • TerraClimate Database (4.6 km spatial resolution): Runoff, Potential evapotranspiration (PET), Minimum temperature (Tmn), Maximum temperature (Tmx), Snow water equivalent (SWE), Actual evapotranspiration (AET), Downward surface shortwave radiation (SRAD).
  • Hydrometeorological drought indices:
    • Standardized Precipitation Evapotranspiration Index (SPEI) at 9, 16, 24, and 36-month scales from the Global SPEI database.
    • Palmer Drought Severity Index (PDSI) from the TerraClimate Database.

Main Results

• Under near-normal hydroclimate conditions (Scenario 1), all models performed well, with LightGBM showing a one-month-ahead forecast skill score of 86.6%, GAM 86.4%, and ProbLSTM 84.7% (using 80% training data).
• When trained with less data (60% historical data) under near-normal conditions, ProbLSTM's skill increased to 87.0%, while LightGBM's performance decreased by approximately 20%.
• Under unprecedented drought conditions (Scenario 2), LightGBM's one-month-ahead skill score significantly dropped to 75.0% (with 80% training data) and further to 63.0% (with 60% training data), exhibiting high uncertainty (NMAE = 11.2%).
• In contrast, ProbLSTM and GAM maintained stable and high predictive performance under unprecedented drought conditions. ProbLSTM achieved the maximum predictive skill score (Skill = 86.2%, NMAE = 3.6%), closely followed by GAM (Skill = 85.3%, NMAE = 3.4%) (using 60% training data).
• For 6-month-ahead forecasts under unprecedented drought, ProbLSTM predictions were closest to observed values, with a minimum Mean Absolute Error (MAE) of 17.7 million cubic meters (Mm³). LightGBM showed error margins 10 times greater than in Scenario 1, with an MAE of 254.2 Mm³.
• The most significant predictors for dam water volume forecasting were previous dam water volume (1-3 months ahead), downward surface shortwave radiation, potential evapotranspiration, soil moisture, runoff, maximum snowmelt, precipitation, prolonged drought conditions (SPEI36), and climate teleconnection indices (PDO, ONI, SOI, Nino1+2).

Contributions

• Developed and optimized hybrid XAI model pipelines (LightGBM, GAM, ProbLSTM) for seasonal dam water volume forecasting using Bayesian optimization (Optuna) and advanced feature selection methods (LASSO, mRMR).
• Provided a comprehensive comparative analysis of model performance under both near-normal and unprecedented drought conditions, highlighting the stability and limitations of different model types in real-world, extreme scenarios.
• Demonstrated the superior and stable predictive capabilities of ProbLSTM and GAM for seasonal water resource forecasting, particularly when faced with unobserved extreme conditions, which is critical for proactive water management.
• Identified and quantified the partial and temporal contributions of a wide range of hydroclimatic and large-scale teleconnection predictors to dam water volume predictability in the Oum Er Rbia watershed, Morocco, using explainable AI techniques (SHAP, Integrated Gradients).
• Addressed the challenge of accurate seasonal forecasting with limited historical data and under conditions of unprecedented changes, offering valuable insights for water resource managers in drought-prone regions.

Funding

• International Climate Cooperation Program (ICCP) from the Quebec Government (reference: 55–2023-BRI-00481).
• University of Quebec at Trois-Rivi`eres (postdoctoral fellowship).
• GEANTech project (collaborative initiative involving OCP Foundation, ESRI, CNRST, and UM6P).

Citation

@article{Houmma2026Seasonal,
  author = {Houmma, Ismaguil Hanadé and Hadri, Abdessamad and Boudhar, Abdelghani and Khalki, El Mahdi El and KARAOUI, Ismail and Oussaoui, Sabir and Samih, Mohamed and Kinnard, Christophe},
  title = {Seasonal forecasting of dam water resources using optimized hybrid models under unprecedented drought conditions},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2025.103091},
  url = {https://doi.org/10.1016/j.ejrh.2025.103091}
}