DEVELOPMENT OF A DELTA-BASED XGBOOST MODEL FOR FORECASTING RESERVOIR STORAGE DYNAMICS

Aliya Aubakirova, Alexandr V. Neftissov, Mukhtar Orazbay, Andrii O. Biloshchytskyi, Tetyana Honcharenko

2026

Closed access Metadata (CSL-JSON)
DOI 10.48081/btouen3.2026/91-113
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Accurately forecasting reservoir water volumes is crucial for climate change adaptation, drought mitigation, and provides a basis for energy planning and management, particularly in continental basins, where hydrological variability poses challenges to traditional modeling. In turn, reduced reservoir inflows limit the water available to hydroelectric power plants for optimal hydroelectric generation. To address this issue, we propose a novel delta-based learning framework using extreme gradient boosting (XGBoost) that forecasts daily volume increments (ΔV) rather than total water volume. The proposed model represents a novel solution where standard machine learning approaches often struggle to capture short-term dynamics due to non-stationary trends in absolute water volume data. The model integrates meteorological and hydrological characteristics of the ERA5 reanalysis lag, is trained on 2019–2022 data, tested on 2023–2025 data, and is adapted to regional specificities. The results demonstrate exceptional performance: the reconstructed volume achieved an R² of 0.983, a Nash–Sutcliffe efficiency (NSE) of 0.983, and a Kling–Gupta efficiency (KGE) of 0.984. Most importantly, the delta-model-based approach reduced the root mean square error (RMSE) by 16 % compared to the baseline direct volume forecasting method, effectively eliminating systematic trend leakage. Performance analysis confirmed the physical dominance of delays in water volume and precipitation integrals. This study confirms that differential dynamics modeling significantly improves the stability and interpretability of forecasts, offering a reproducible and reliable tool. Keywords: reservoir forecasting, delta-based learning, hydrological machine learning, XGBoost, ERA5 reanalysis, multi-horizon prediction, remote sensing, predictive modeling.