Attention-driven deep learning models for multivariate time series forecasting of reservoir water levels

Accurate reservoir water level forecasting is crucial for effective water resource management, flood control, and irrigation planning. This study investigates the application of deep learning methods to forecast daily water elevation in the Nesa Dam, located in southeastern Iran. A 15-year dataset comprising daily hydrometeorological variables - including rainfall, temperature, evaporation, inflow, and outflow - was used to train and evaluate three models: CNN + BiLSTM + Attention, Encoder-Decoder LSTM with Attention, and ConvLSTM2D.Using a sliding window approach, the models were trained on 80% of the data and tested on the remaining 20%. Performance was evaluated using RMSE, MAE, and R². The Encoder-Decoder LSTM with Attention model achieved the best results, with the lowest prediction errors and highest generalization performance. The CNN-based model showed moderate accuracy, while ConvLSTM2D exhibited noisy outputs and limited predictive capability. The study demonstrates that attention-based architectures are highly effective in modelling temporal dependencies in hydrological time series. The novelty lies in the comparative analysis of these models under identical experimental conditions using real-world data. The findings offer practical insights for developing intelligent forecasting systems in water resource management.