Using Temporal Convolutional Neural Networks to Project Streamflow
A novel temporal convolutional neural network (TCNN) is developed for producing a local streamflow time series given precipitation as input. This machine learning model demonstrates superior performance to competing machine learning architectures when training is performed on a per-basin basis. By analyzing its sensitivity to input features and input time window size, we can distinguish different dominant physical processes for the streamflow dynamics in each basin, and how streamflow in those basins are expected to respond to climate change. The difference in the dominant physical process is also reflected in future projection.
Streamflow has been found to be notoriously difficult to calculate in arbitrary watersheds because of strong sensitivities within process-based hydrologic models. However, machine learning methods have been developed over the past several years that have demonstrated superior performance on streamflow prediction. With a growing need for streamflow information in the future, there has been a pressing need to develop systems that can be used for future streamflow projection as well.
Machine learning models have been demonstrated to be generally superior to process-based hydrologic models for streamflow dynamics, however, we still expect that variations in the architecture of model can lead to incremental improvements in the quality of these models for streamflow projection. TCNNs are advantageous because they are generally faster to train than LSTMs while demonstrating competitive performance. We have demonstrated that these models can further be used for streamflow projection under a non-stationary climate, where precipitation and temperature are above historical values and allow us to identify clear differences in the response of particular basins because of their individual characteristics.