Deep learning for streamflow forecasting in semi-arid basins

Accurate streamflow forecasting is crucial for effective water resources management, particularly in semi-arid regions increasingly impacted by climate change. This study evaluated the performance of deep learning models for streamflow forecasting in two catchments in eastern Spain. The models were trained on historical data using a one-step-ahead forecasting approach and evaluated through temporal cross-validation. A recursive multi-step forecasting strategy was subsequently used to assess predictive performance across different forecasting horizons. The long short-term memory (LSTM) models generally outperformed the multilayer perceptron (MLP) models due to their ability to capture temporal dependencies, although they exhibited high sensitivity to the length of training data and model calibration. The MLP models performed better with simple preprocessing, whereas the LSTM models benefited from combining temporal features with deseasonalization techniques. The optimal configuration for each catchment consistently delivered robust performance and reasonable predictions across various forecasting horizons. This study highlights the potential of neural network models for streamflow forecasting and provides practical guidance for implementing deep learning models in semi-arid basins, thereby contributing to improved drought risk assessment and water resources management.