Journal
JOURNAL OF HYDROLOGY
Volume 628, Issue -, Pages -Publisher
ELSEVIER
DOI: 10.1016/j.jhydrol.2023.130504
Keywords
Streamflow prediction; Artificial Intelligence; Scale independent performance metrics
Ask authors/readers for more resources
Machine learning algorithms have shown better performance in streamflow prediction compared to traditional hydrological models. In this study, researchers proposed a methodology to test and benchmark ML algorithms using artificial data generated by physically-based hydrological models. They found that deep learning algorithms can correctly identify the relationship between streamflow and rainfall in certain conditions, but fail to outperform traditional prediction methods in other scenarios.
Machine learning (ML) algorithms have produced remarkable advances in streamflow prediction, exceeding the performance of calibrated conceptual and physics-based hydrological models that have been developed over many decades. ML algorithms seem to overcome the issue of errors known to be present in rainfall and streamflow estimates that have hindered the performance of hydrological models for decades. In this paper, we propose a methodology for testing and benchmarking ML algorithms using artificial data generated by physically-based hydrological models. Our approach makes it possible to design controlled numerical experiments that can improve our understanding of this new generation of black-box models. We conducted a diagnostics study to demonstrate our methodology in which we attempted to determine if ML algorithms can identify a function relating streamflow and rainfall. This exercise combined the implementation of the distributed hillslope-link hydrological model (HLM) on a 4,385 km2 basin driven by precipitation fields created using the stochastic storm transposition (SST) framework, and an advanced deep learning algorithm based on gated recurrent unit (GRU)-Attention neural networks. The data generated allowed us to create prediction scenarios that are equivalent to the hindcast and real-time forecast problems. We proposed a set of scale-independent performance metrics to evaluate the results of our experiment and found that the GRU can correctly identify a predictive function for all analyzed locations in the river network. We concluded that under the circumstances tested in this study, deep learning can identify the transformation function when trained in Hindcast Mode, making it a powerful tool to determine the streamflow response of a basin to predetermined rainfall scenarios. However, it fails to significantly outperform the predictions of temporal persistence when tested in Forecast Mode.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available