Reliable short- to medium-term river discharge forecasting is a relevant problem,
as confirmed by increasingly frequent flood events. Discharge forecasting
is challenging due to the complex temporal and spatial dynamics between
river discharges and meteorological variables such as precipitation.
The Slovenian Environment Agency (ARSO) currently uses the Hydrological
Forecasting System (HPS), which addresses the problem with traditional
methods based on physical equations. In this work, we propose an alternative
approach called SIREN (Synoptic Information-based River Discharge
Estimation Network), that is based on a modern architecture of deep neural
networks, which are known for their effectiveness in modeling complex patterns.
On a real-world dataset, the SIREN model was thoroughly analyzed
and compared with HPS and with the HIDRA3river model, which is based
on the HIDRA3 architecture, a deep neural network for sea level forecasting.
The results showed that the SIREN model is more effective than the
reference methods. On selected metrics, it achieves up to 62.8% better performance
than HPS and up to 30.7% better performance than HIDRA3river,
thus representing the new state of the art in river discharge forecasting for
Slovenian rivers.
|
