Flood prediction through hydrological modeling of rainfall using Conv1D-SBiGRU algorithm and RDI estimation: A hybrid approach

Abstract

Time series prediction of natural calamities is effectively solved with deep neural networks due to their ability to automatically assimilate the temporal linkages in time series data. This research develops a hybrid stacked deep learning with one-dimensional Convolution–Stacked Bidirectional Gated Recurrent Unit (Conv1D-SBiGRU) algorithm, unifying the predictive advantages of one-dimensional Convolution (Conv1D) and Bidirectional Gated Recurrent Unit (BiGRU) using hydro-meteorological and atmospheric data to build and evaluate a flood prediction model in forecasting the phenomenon of forthcoming flood events. The one-dimensional Convolution model effectively obtains valuable information and learns the time series cognitive representation. The stacked BiGRU model efficiently identifies and models the data sequence with temporal dependencies due to their ability to learn from past and future moments. The developed predictive model uses statistically significant predicted rainfall value to estimate the daily Relative Departure Index (RDI) which is used to predict floods. The proposed work was trained and evaluated for predicting floods on the real-world data of Alappuzha district, Kerala, India. The findings demonstrate the preeminence of the Conv1D-SBiGRU-based flood model with around 33% reduced MAE and RMSE and 9% improved R² over the benchmark and some hybrid techniques. The outcomes showed the efficiency of Conv1D-SBiGRU in precisely forecasting floods during extreme weather events with an accuracy of 98.6%.