Applying Deep Learning Models for Catchment Scale Streamflow Prediction

Abstract:

Accurate streamflow prediction is crucial for water resources management and flood control. From literature, both physically based and data-driven models have been applied for streamflow forecasting. Physical-based models require knowledge of physical processes and high computational power; while data-driven models, such as artificial neural networks (ANN) work by learning the underlying relationships between inputs and outputs without explicitly referencing the physical process. Compared to the conventional ANN, Long Short-Term Memory (LSTM) has feedback connections, which makes it effective in time series prediction. However, the convolutional neural network (CNN) could better process high dimensional data than LSTM by utilizing filters to extract local connections and features to compose large-scale patterns.

In this study, we developed a CNN model for streamflow forecasting in Little Washita Watershed in Oklahoma. Ten years (2010-2019) of data, with nine years (2010-2018) for calibration and one year (2019) for validation, including rainfall, soil moisture, past streamflow, were collected for the experiment. The simulation results of the CNN were compared with that of LSTM, multi-layer feedforward neural network (MFN) and ParFlow-CLM, a physical-based hydrological model. Validation for three stream gage locations within the watershed shows CNN had the best performance, with R² ranging from 0.976 to 0.995 and RMSE ranging from 0.13 m³/s to 0.51 m³/s. The fully connected layer in the CNN model acts like a classification filter to classify the flow pattern effectively based on different input features. The ParFlow-CLM model reproduced the distinct patterns in the dry and wet seasons, but overestimated the peak flows. The MFN and LSTM estimated the peak flow of the validation dataset correctly, but missed the peak flow for the calibration data. Based on the results, the CNN model, which can provide reliable streamflow prediction in both dry and wet seasons, is the most ideal for flooding warning of the watershed.