Global climate change and increasing water demand pose serious challenges for sustainable water resource management, particularly in large metropolitan areas such as Istanbul. Accurate forecasting of dam fill rates is therefore critical for proactive water management and drought mitigation strategies. In this study, we propose a Residual-Based Hybrid BiLSTM–XGBoost (RBH-BiLSTM-XGB) model to forecast dam fill rates using multivariate time series data, including daily precipitation, water consumption, and historical reservoir levels obtained from the Istanbul Metropolitan Municipality Open Data Portal. The proposed approach combines the temporal dependency learning capability of Bidirectional Long Short-Term Memory (BiLSTM) networks with the residual error modeling strength of XGBoost, enabling the correction of systematic prediction errors produced by deep learning models. The proposed hybrid model is evaluated against standalone LSTM, BiLSTM, XGBoost, and Prophet models under different training–testing split scenarios (70–30%, 80–20%, and 90–10%). Model performance is assessed using RMSE, MAE, MAPE, and R² metrics. Experimental results demonstrate that the RBH-BiLSTM-XGB model consistently outperforms all benchmark models, achieving the lowest RMSE (0.0059), MAPE (0.66–0.86%), and the highest explanatory power (R² ≈ 0.9994). While deep learning models effectively capture long-term temporal dependencies, tree-based models such as XGBoost are shown to be effective in learning residual structures that deep networks fail to model. In contrast, the Prophet model exhibits poor performance due to its additive structure and limited capacity to represent complex multivariate interactions. The findings highlight the effectiveness of residual-based hybrid modeling for dam fill rate forecasting and demonstrate the potential of integrating deep learning and ensemble learning approaches to support data-driven and sustainable water resource management.
Hikmet Canlı (Sun,) studied this question.