Please login first
Development of Storm Surge Prediction Model Using Artificial Neural Network in the Southern Region of Bangladesh
* 1 , 2 , 3
1  MS student, Department of Disaster Science and Climate Resilience, University of Dhaka
2  Professor, Department of Disaster Science and Climate Resilience, University of Dhaka
3  Directorate of Groundwater Hydrology, Bangladesh Water Development Board
Academic Editor: Wataru Takeuchi

Abstract:

Climate change and the alarming rate of sea level rise are pausing more threats to the exposed coastal region of Bangladesh from destructive storm surges, as well as fueling the demand for speedier and more reliable surge forecast systems that predict the storm surge with sufficient lead time. However, the currently used physics-based numerical models are computationally costly and time-consuming for surge forecasting and early warning. To address this situation, the current study presents two-layer feed-forward ANN models for 24, 18, 12, and 6-hour lead times to quickly predict the surge height along the exposed coast of Bangladesh using the time series of tropical cyclone parameters and tidal data. Seven historical tropical cyclones from 1995 to 2022, including Sidr (2007), Aila (2009), Roanu (2016), and Sitrang (2022), were chosen to construct the input layer for this study based on the availability of data. The study used tidal level (provided by BIWTA and UHSLC), longitude and latitude of the cyclone eye, central atmospheric pressure, pressure gradient, highest wind speed, and distance from the cyclone eye to the area of interest (provided by NOAA and IMD) as input components. The models were developed through a neural fitting tool in MATLAB, and the best-fitted models were determined using the optimal combination of input layer parameters and the hidden layer, or the number of hidden neurons. In the application of the models to Cox's Bazar, Chittagong, Hiron Point, and Charmontaj, it was found that the best-performing models for a 24, 18, 12, and 6-hour lead time had the best-fit sets of hidden neurons of 30, 50, 20, and 70, respectively. The RMSE of the models spans between 0.121 (for a 12-hour lead time) and 0.151 (for a 24-hour lead time), indicating high precision and accuracy. The models are also rapid, predicting the water level in less than a minute. In short, the proposed method can be adapted to develop reliable models for forecasting surge levels at any other coastal site in Bangladesh.

Keywords: Storm surge; Forecast model; Tidal surge prediction; Artificial neural network; Data driven model
Comments on this paper
Adeline Smith
best explanation so far! thank you. i have tried it in R using the neuralnet function with your dataset. even though i get the same coefficients with the log regression the weights and bias using the ANN are not the same. they are much lower. any idea why?
P/s: Despite being a free version, Geometry Dash Lite still manages to capture the essence of the core game.

Adeline Smith
best explanation so far! thank you. i have tried it in R using the neuralnet function with your dataset. even though i get the same coefficients with the log regression the weights and bias using the ANN are not the same. they are much lower. any idea why?
P/s: Despite being a free version, Geometry Dash Lite still manages to capture the essence of the core game.

otis jame
Players Sprunki can mix and match these sound loops by dragging and dropping icons onto the characters, creating custom musical tracks. As more sounds are added, the composition becomes richer and more intricate, allowing players to experiment with different musical styles.




 
 
Top