Accurate reservoir water level forecasting is essential for water resource management, flood risk mitigation, and hydropower operation. However, it remains challenging due to pronounced geographical heterogeneity and complex multi-scale temporal dynamics. Existing deep-learning approaches typically overlook explicit geographical and climatic conditioning. They struggle to capture temporal dependencies across multiple time scales. They also exhibit limited transferability across reservoirs with similar hydrological characteristics. To address these limitations, this paper proposes GeoFormer, a geography-aware adaptive Transformer framework designed for reservoir water level forecasting across diverse geographical contexts. GeoFormer integrates three key innovations. First, a Geography-Aware Embedding Module conditions temporal representations on geographical location, climate regimes, and reservoir attributes. Second, an Adaptive Multi-Scale Temporal Fusion mechanism dynamically aggregates information across daily, weekly, and monthly temporal resolutions. Third, a Cross-Reservoir Knowledge Transfer strategy enables effective knowledge sharing among hydrologically similar reservoirs. Extensive experiments on six reservoirs distributed across multiple continents and climate zones demonstrate that GeoFormer consistently outperforms state-of-the-art baselines, including iTransformer, DLinear, and Informer. The model achieves average reductions of 23.7% in RMSE, 19.4% in MAE, and 15.8% in MAPE, while maintaining strong robustness and generalization across geographically heterogeneous hydrological systems.
Wu et al. (Sat,) studied this question.