“We pre-trained the model at several continental scales using basins with historical monitoring data,” said corresponding author and professor Ouyang Chaojun of the CAS Institute of Mountain Hazards and Environment. Ta. He explained that this allows flow predictions in basins where there are no flow records.
The researchers write in their paper: “Our proposed model performs better in transregional flow prediction tasks compared to other machine learning models and classical hydrological models. Achieved cutting-edge performance.
“River flow and flood prediction remains one of the long-standing challenges in hydrology.”
Many predictive models require high-quality historical data, highlighting the huge challenge of developing reliable flow forecasts for thousands of catchments without access to physical parameters or historical data. “There is,” the researchers wrote.
Recent studies have also focused on forecasting specific regions, using local data that “precludes a universal assessment of global flow forecasts,” the researchers said.
To achieve this, the researchers proposed a model that uses only meteorological forcing inputs such as rainfall, temperature, and static land attributes.
Static attributes such as soil characteristics “can be obtained from globally available satellite data,” the researchers said.
02:06
Chinese residents battle record storms
Chinese residents battle record storms
“For the first time, multiple hydrological AI models have been trained to provide comparative analysis on a global scale,” the team wrote.
Within their model, spatial attributes and climate characteristics over time are treated separately. This differs from other models that use aggregate indices, “resulting in greater bias in simulations and predictions,” Ouyang said.
“Compared to other models, ED-DLSTM shows superior predictive ability.”
The model pre-trained in the US was the most efficient, with nearly 77% of catchments having an NSE greater than 0.
The researchers said their tests verified that “the model is able to learn universal hydrological behavior with different training sets.”
“This study demonstrates the potential of deep learning methods to overcome the ubiquitous lack of hydrological information and deficiencies in the structure and parameterization of physical models.”
