Leveraging Flow Duration Information to Improve Streamflow Prediction in Data-sparse Regions with Deep Learning Models

Abstract:

While time series deep learning (DL) has recently demonstrated stellar performance in streamflow prediction, those models are hungry for data and face difficulties when applied to large, contiguous regions with sparse or no gauges. However, even in these regions, flow distribution information such as the flow duration curve (FDC) may be available from various sources, or will be available with future satellite missions. Here we demonstrate that we could make use of scattered FDC or short instances of streamflow records to improve the predictions in poorly-gauged regions. We trained long-short term memory models (LSTM) with convolutional neural network (CNN) kernels that can integrate the FDC. The results show that the CNN kernels can extract useful characteristics from FDC. The integration of FDC, with or without generic catchment attributes, could improve streamflow predictions in poorly-gauged regions. The proposed CNN-LSTM framework could achieve the migration of FDC information to ungauged basins and maximally leverage both flow distribution and daily discharge data, which is implausible for traditional methods. It can be argued that this model is of highly practical value to the streamflow predictions in data-sparse regions.