A Comparative Flood Forecasting Analysis Using Different Deep Learning Techniques for Baitarani Basin in India

Abstract This paper explores different Deep Learning (DL) models for flood forecasting at Baitarani basin in India. Here, three different modeling techniques namely Long Short Term Memory (LSTM), Convolution Neural Network (CNN) and Encoder-Decoder (E-D) have been chosen for flood forecasting. Analysis is performed at two gauging sites of Baitarani basin namely Champua and Anandpur, taking 30 years (1991–2020) data of discharge, rainfall and water level as input. Different statistical indices are used for model evaluation. From the study it was found that CNN model efficiency is more than 91% for a lead time of 1-day and 86% for a lead time of 2day at Champua gauging station but as moving towards downstream gauging site of Anandpur, LSTM is performing better that CNN and E-D model. LSTM model efficiency is more than 88% for a lead time of 1-day and 80% for a lead time of 2-day at Anandpur gauging station.