Enhanced Forecast Skill in Dendritic Catchment Through Streamflow Imputation: Application of Physics-Informed Graph Neural Network

Gaps and inconsistencies in streamflow records limit the application of deep learning models for flood forecasting. Here, we develop a novel framework that first imputes missing streamflow records and use the reconstructed record for streamflow forecasts at 25-gauge stations in the Godavari River Basin, India. The imputation model reconstructs missing streamflow records using a Physics-Informed graph neural network (GNN)-bidirectional Long Short-Term Memory (LSTM) architecture during1980-2020. The resulting gap-filled streamflow is subsequently used as input to the probabilistic forecast model. The forecast framework, designed to provide 14-day lead forecasts , combines a GNN with an encoder-decoder LSTM and an attention mechanism. The imputation model successfully captures peak flows, achieving median Nash Sutcliffe Efficiency (NSE) and Kling Gupta Efficiency (KGE) of 0.99 and 0.96, respectively during multi-year gaps. The forecast model trained with imputed streamflow (FC-Imputed) achieves median NSE and KGE values of 0.73 and 0.71 , respectively, across all stations and lead times. Compared with forecasts based on observed streamflow records (FC-Observed), FC-Imputed improves the median NSE and KGE by 52% and 30% across all lead times. Moreover, we find significant improvement in forecast model performance during peak flows (Q>Q 95 ), where the median Continuous Ranked Probability Skill Score (CRPSS) increased from 0.69 to 0.90 for stations with less than a 10% missing record at 1–3-day lead. Further, we find that a consistent antecedent streamflow record is crucial for improving forecast performance at short-lead times. The proposed framework is transferable and has the potential to improve short- to medium-term streamflow forecasts in data-scarce river basins.