Multi Day Ahead Flood Predictionin South Asian Tropical Zone Using Deep Learning

Abstract

A reliable and accurate flood forecasting procedure is a critical need due to the hazardous nature of the disaster. Researchers are increasingly favoring innovative approaches with enhanced accuracy, such as machine learning models, over traditional methods for this task. However, lack of such studies regarding South Asian tropical region, which has its own climate characteristics, was unidentified as a major issue. This research delves into the viability of employing ANN, LSTM, BLSTM, ConvLSTM2D and Transformer models for multi-day ahead flood simulation. One-day, two-days and three-days were selected as lead times for the task considering the lower reaches of the Mahaweli catchment in Sri Lanka, which is mostly affected by the Northeast Monsoon. The prediction capability of extreme stream flows was also of interest. Observed rainfall data from three nearby rain gauges, along with historical discharges of the target river gauge, serve as input features for the models. The ANN model showed the worst performance, having the mean NSE of 0.67. An improved performance was observed from the Transformer compared to the LSTM based models, especially in multiple day ahead forecasting scenarios. For all the models, the forecasting capability of extreme water levels drops down drastically when the prediction lead time is increased.

Behavior of LSTM and Transformer Deep Learning Models in Flood Simulation Considering South Asian Tropical Climate

“…Historical data of the three rain gauges and the target river gauge were used as inputs to forecast the following day water level using the sliding window method. Hyperparameters for the models were kept same as in our previous paper (Madhushanka et al, 2024), which were determined by trial-and-error approach. The Transformer architecture employed has been slightly modi ed from the original implementation presented in (Vaswani et al, 2017).…”

“…As outlined in our previous paper (Madhushanka et al, 2024), the analysis was conducted using rainfall data from the designated rain gauge stations along with the water level at Manampitiya, based on the Pearson correlation coe cient calculated among the four stations. Because upstream river gauges belong to a region with different geographic and climatic conditions, they were considered not to be used.…”

“…When considering the Tropical regions, especially areas in South Asia, lack of studies regarding deep learning-based ood simulation is an issue. This paper is related to our study (Madhushanka et al, 2024) which focused the lower reach of the Mahaweli catchment, Sri Lanka. In this paper, apart from the forecasting capabilities of daily stream ows, the effects of different input features on the output are thoroughly investigated.…”