Long-term streamflow forecasting is a critical step when planning and managing water resources. Advanced techniques in deep learning have been proposed for forecasting streamflow. Applying these methods in long-term streamflow prediction is an issue that has received less attention. Four models, including Adaptive Neuro-Fuzzy Inference Systems (ANFIS), Artificial Neural Networks (ANN), Bidirectional Long-Short Term Memory (BiLSTM), and hybrid Convolutional Neural Network (CNN)-Gated Recurrent Unit (GRU)-LSTM, are applied to forecast the long-term daily streamflow of the Colorado River in the U.S. The proper time lag for input series creation is determined using partial autocorrelation. 60% of the data (1921–1981) is used for training, whereas 40% (1981–2021) is used to evaluate the model’s performance. The results of the studied models are assessed by Using four indices: the Mean Absolute Error (MAE), the Normalized Root Mean Square Error (NRMSE), the Correlation Coefficient (r), and the Nash–Sutcliffe Coefficient (ENS). As a result of the testing step, the ANFIS model with NRMSE = 0.118, MAE = 26.16 (m3/s), r = 0.966, and ENS = 0.933 was more accurate than other studied models. Despite their complexity, the BiLSTM and CNN-GRU-LSTM models did not outperform the others. Comparing these models to ANN and ANFIS, it is evident that their performance is not superior.

Keywords