Data Assimilation and Autoregression for Using Near-Real-Time Streamflow Observations in Long Short-Term Memory Networks

How can we take advantage of near-real time streamflow observations to make the best streamflow forecasts into the future?

This paper details how we incorporate recent streamflow observations into our deep learning hydrologic models to improve our forecasts using prediction errors in the recent past to adjust the model’s internal hydrologic states, providing it with a more accurate understanding of current conditions in the basin as it begins to forecast. In the paper we benchmark and discuss the tradeoffs of a few approaches for incorporating recent streamflow observations into deep learning hydrologic models. The gradient based approach we use allows HydroForecast to continue making predictions when live gauge feeds go down while taking full advantage of near-real time observations when they are available.

Read the paper here.

Nearing, G. S., Klotz, D., Frame, J. M., Gauch, M., Gilon, O., Kratzert, F., Sampson, A. K., Shalev, G., and Nevo, S.: Technical note: Data assimilation and autoregression for using near-real-time streamflow observations in long short-term memory networks, Hydrol. Earth Syst. Sci., 26, 5493–5513,, 2022.