Improving Long-Term Flood Forecasting Accuracy Using Ensemble Deep Learning Models and an Attention MechanismSource: Journal of Hydrologic Engineering:;2024:;Volume ( 029 ):;issue: 006::page 04024042-1Author:Marjan Kordani
,
Mohammad Reza Nikoo
,
Mahmood Fooladi
,
Iman Ahmadianfar
,
Rouzbeh Nazari
,
Amir H. Gandomi
DOI: 10.1061/JHYEFF.HEENG-6262Publisher: American Society of Civil Engineers
Abstract: Floods, as major natural disasters, cause massive property destruction and death. Understanding the occurrence time of this event by advance notice helps consider operational flood prevention systems and platforms. Precise flood forecasting provides a suitable time for policymakers and the public to consider helpful responses to this event. This study introduces an innovative methodology to enhance the precision of long-term flood predictions by employing a multistep forecasting approach. Our approach leverages historical time-series data on precipitation and streamflow to train an autoencoder algorithm. The primary objective is to develop advanced forecasting models to predict 12-step-weekly ahead flood occurrences during the critical April to July period from 2019 to 2021 within the DuPage River basin, Illinois, USA. In order to achieve this goal, we explore three deep learning techniques: bidirectional long short-term memory (BI-LSTM), ensemble long short-term memory (E-LSTM), and ensemble long short-term memory-gated recurrent unit (E-LSTM-GRU). Then, we integrate an attention mechanism (AM) that utilizes dynamic fusion techniques to emphasize the salient features of ensemble models. A dedicated fusion model is developed for each forecasting stage, effectively consolidating the predictions from various deep-learning models. Additionally, two traditional machine learning techniques, namely MLP and SVM models, are used to compare and justify the efficiency of applied deep learning models. The performance evaluation of our approach using statistical error metrics, including coefficient of determination (R2), normalized root mean square error, normalized mean absolute error, mean absolute percentage error, Nash–Sutcliffe efficiency, Kling–Gupta efficiency, and percent bias, for the 12th step prediction reveals impressive results, with average values of 0.976, 2.393, 1.892, 20.956, 0.967, 0.923, and 2.307, respectively. These findings underscore the capability of our proposed models to significantly reduce uncertainty in flood forecasting, thus enhancing the reliability and accuracy of future predictions.
|
Collections
Show full item record
contributor author | Marjan Kordani | |
contributor author | Mohammad Reza Nikoo | |
contributor author | Mahmood Fooladi | |
contributor author | Iman Ahmadianfar | |
contributor author | Rouzbeh Nazari | |
contributor author | Amir H. Gandomi | |
date accessioned | 2025-04-20T10:27:42Z | |
date available | 2025-04-20T10:27:42Z | |
date copyright | 9/14/2024 12:00:00 AM | |
date issued | 2024 | |
identifier other | JHYEFF.HEENG-6262.pdf | |
identifier uri | http://yetl.yabesh.ir/yetl1/handle/yetl/4304767 | |
description abstract | Floods, as major natural disasters, cause massive property destruction and death. Understanding the occurrence time of this event by advance notice helps consider operational flood prevention systems and platforms. Precise flood forecasting provides a suitable time for policymakers and the public to consider helpful responses to this event. This study introduces an innovative methodology to enhance the precision of long-term flood predictions by employing a multistep forecasting approach. Our approach leverages historical time-series data on precipitation and streamflow to train an autoencoder algorithm. The primary objective is to develop advanced forecasting models to predict 12-step-weekly ahead flood occurrences during the critical April to July period from 2019 to 2021 within the DuPage River basin, Illinois, USA. In order to achieve this goal, we explore three deep learning techniques: bidirectional long short-term memory (BI-LSTM), ensemble long short-term memory (E-LSTM), and ensemble long short-term memory-gated recurrent unit (E-LSTM-GRU). Then, we integrate an attention mechanism (AM) that utilizes dynamic fusion techniques to emphasize the salient features of ensemble models. A dedicated fusion model is developed for each forecasting stage, effectively consolidating the predictions from various deep-learning models. Additionally, two traditional machine learning techniques, namely MLP and SVM models, are used to compare and justify the efficiency of applied deep learning models. The performance evaluation of our approach using statistical error metrics, including coefficient of determination (R2), normalized root mean square error, normalized mean absolute error, mean absolute percentage error, Nash–Sutcliffe efficiency, Kling–Gupta efficiency, and percent bias, for the 12th step prediction reveals impressive results, with average values of 0.976, 2.393, 1.892, 20.956, 0.967, 0.923, and 2.307, respectively. These findings underscore the capability of our proposed models to significantly reduce uncertainty in flood forecasting, thus enhancing the reliability and accuracy of future predictions. | |
publisher | American Society of Civil Engineers | |
title | Improving Long-Term Flood Forecasting Accuracy Using Ensemble Deep Learning Models and an Attention Mechanism | |
type | Journal Article | |
journal volume | 29 | |
journal issue | 6 | |
journal title | Journal of Hydrologic Engineering | |
identifier doi | 10.1061/JHYEFF.HEENG-6262 | |
journal fristpage | 04024042-1 | |
journal lastpage | 04024042-19 | |
page | 19 | |
tree | Journal of Hydrologic Engineering:;2024:;Volume ( 029 ):;issue: 006 | |
contenttype | Fulltext |