Authors: William Currier – NOAA Physical Sciences Laboratory
Title: Does Deep Learning Absorb Ensemble Forecast Post-Processing Benefits for Probabilistic Streamflow Prediction?
Presentation Type:
Abstract: Operational hydrologic forecasting systems typically treat meteorological post-processing and hydrologic modeling as sequential, independent steps: raw ensemble forecast output is first statistically corrected with observed meteorology, then routed through a calibrated hydrologic model. Recent advances in deep learning for hydrology have demonstrated that encoder-decoder architectures can achieve strong forecast skill from raw numerical weather prediction inputs. However, whether this extends beyond deterministic point forecasting to the full probabilistic setting—and whether it renders explicit post-processing necessary—remains an open question.
We’ve designed a controlled experiment that evaluates the marginal value of explicit GEFS ensemble post-processing when the hydrologic model itself can learn forecast error characteristics. Using 20 years of GEFS reforecasts across headwater basins, we compare per-basin calibrated SAC-SMA driven by both raw and post-processed forcings against a regionally trained encoder-decoder LSTM with a mixture density output head, evaluated at NOAA’s operational priority basins. Each GEFS ensemble member is processed through a separate decoder forward pass, producing member-specific predictive distributions that preserve the separation between meteorological forcing uncertainty and hydrologic model uncertainty. Our experimental framework evaluates using proper probabilistic scoring rules across lead times, flow regimes, and event magnitudes to understand where and when explicit post-processing still matters.