Utilizing Deep Learning Models to Predict Streamflow

This study employs convolutional neural network (CNN), long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM), and gated recurrent unit (GRU) deep learning models to simulate daily streamflow using precipitation data. Two approaches were explored: one without dimension reduction and another incorporating dimensionality reduction technique. Principal component analysis (PCA) was employed for dimensionality reduction, and partial autocorrelation function (PACF) was used to determine time lags. An augmented Dickey–Fuller (ADF) test was utilized to ascertain the stationarity of the data, ensuring optimal model performance. The data were normalized and then partitioned into features and target variables, before being split into training, validation, and test sets. The developed models were tested for their performance, robustness, and stability at three locations along the Neuse River, which is in the Neuse River Basin, North Carolina, USA, covering an area of about 14,500 km2. Furthermore, the model’s performance was tested during peak flood events to assess their ability to capture the temporal resolution of streamflow. The results revealed that the CNN model could capture the variability in daily streamflow prediction, as evidenced by excellent statistical measures, including mean absolute error, root mean square error, and Nush–Sutcliffe efficiency. The study also found that incorporating dimensionality reduction significantly improved model performance.