A hybrid data assimilation and machine learning approach for enhancing operational forecasting in 2D hydrodynamic models

ABSTRACT Data assimilation (DA) integrates observations to enhance model predictions, but its benefits can rapidly decay during forecast for models heavily influenced by boundary conditions. To extend forecast accuracy, we combine machine learning (ML) with DA in a hybrid approach. We utilize ML to create synthetic point observations, which are assimilated into the model and thus propagated to other positions and model variables. In a case study, we apply the approach to a tidal estuary hydrodynamic model using a long short-term memory (LSTM) network for ML forecasts and the Ensemble Kalman Filter (EnKF) for assimilation. Hereby, physical consistency and spatial representation of water levels in the estuary are maintained. Our results from 100 forecast runs show significant improvements: over 40% accuracy increase in forecasts up to 4 h ahead, and a >5% error reduction at 9 h compared with a traditional numerical model without assimilation.