Tested near Chicago, the model predicts water levels across 4.2 million 30-metre cells up to one day in advance, around 150 times faster than a hydrodynamic model.
During a flood, crisis managers need to know where the water is, how high, and how it will evolve in the following hours. But they only have a few water level stations, and the high-resolution hydrodynamic models are too slow to be restarted with each new measurement. Researchers present C-STRIDE, an observation-driven digital twin that converts short histories of a few stations, combined with terrain and rainfall, into water depth maps over an entire basin, with a projection of up to 24 hours.
Trained on simulations of a calibrated 2D hydrodynamic model, it dispenses with the need for data assimilation. In the Des Plaines River basin, near Chicago, six stations are enough to feed forecasts on 4.2 million 30-meter grids. The relief is the most useful information, the rain prevents errors from drifting with the horizon; together, they reduce the error by about 40% compared to measurements alone. With known future rainfall, the error remains close to 15% at one day, compared to nearly 40% without.
Fed with real measurements rather than simulated data, the model adjusts its forecasts to the hydrographs observed for three out of six stations, without retraining. The authors acknowledge that it still needs to be validated in real time with operational rainfall forecasts.
Read the full article on the original source
Go to the source →