Forecasting flood inundations in the dam-regulated Mahanadi River delta using integrated hydrologic-hydrodynamic-deep learning model

Abstract

The efficacy of a deep learning error-updating model in predicting the hydrological model-simulated errors influenced by reservoir regulation is assessed. Two daily discharge forecasting model frameworks without (Case I) and with (Case II) error-updating of the discharge forecasts at a downstream location are developed. The best discharge forecasts are forced as inputs to a hydrodynamic model to simulate the forecasted flood inundations in the downstream region. The findings reveals that the discharge forecasts with the forecasted releases from the reservoir as upstream inflow boundary, post-error updating at the delta head (Case II) outperforms Case I with an $NSE$ value of 0.83–0.94 at 1–5 days lead times. Moreover, this model (Case II) captures the flood peaks with the least error and narrowest uncertainty bands. Further, with a 49–52 % coincidence of observed and simulated flood inundation extent, the hydrodynamic model simulates the inundation forecasts with reasonable accuracy up to 5-days lead times.