Multiple-Day-Ahead Flood Prediction in the South Asian Tropical Zone Using Deep Learning

A reliable and accurate flood forecasting procedure is crucial due to the hazardous nature of such disasters. Despite the growing interest in machine learning models over traditional methods for enhanced accuracy, there is a notable gap in guidelines for selecting the best-suited learning algorithm for flood forecasting. Furthermore, there is a lack of deep learning–based flood simulation studies in the South Asian Tropical Zone. This research addresses these gaps by investigating the viability of artificial neural network (ANN), long short-term memory (LSTM), bidirectional LSTM (BLSTM), two-dimensional (2D) convolutional LSTM (ConvLSTM2D), and transformer models for multiple-day-ahead flood simulation. A forecasting window of 3 days was selected for the task, focusing on the lower reaches of the Mahaweli catchment in Sri Lanka, which is heavily affected by the Northeast Monsoon. Observed rainfall data from three nearby rain gauges and historical discharges from the target river gauge serve as input features for the models. Unlike previous studies that used limited data sets, this study utilizes daily data spanning 28 years in order to examine the behavior of the transformer handling an extensive hydrological data set. The study finds that the ANN model performed the worst, with a mean Nash-Sutcliffe efficiency (NSE) of 0.67, while the transformer model showed superior performance, especially in multiday forecasts, with a mean NSE of 0.72 and a mean root-mean square error of 32.52, showcasing the effectiveness of handling this extensive data set.