Flood is difficult to predict due to its extreme runoff values, short duration and complex generation mechanism. In order to reduce the negative effects of flood disasters, researchers try to forecast flood by utilizing deep learning technology. Essentially, historical flood data can be regarded as sequential data with sets of flood factors. Facing challenges brought by inherent characteristics of flood forecasting, this paper proposes a dual attention embedding network, i.e., DA-Net, to achieve accurate prediction results. The proposed attention mechanism not only embeds a convolution self-attention module (CSA) on Temporal Convolutional Network (TCN) for description of local context information, but also constructs a Temporal-related Feature Attention (TFA) Module to assign time-varying weights for different features in a global sense. Specifically, CSA offers additional and local context information to help predict extreme runoff values even within a small period, meanwhile TFA improves global modeling capability of TCN for construction of data-driven generation mechanism in our method. Experiments on Changhua and Tunxi watershed dataset show the proposed method achieves superior prediction performance than current deep learning based methods.

DA-Net: Dual Attention Network for Flood Forecasting

Castiglione A.;Narducci F.;
2023-01-01

Abstract

Flood is difficult to predict due to its extreme runoff values, short duration and complex generation mechanism. In order to reduce the negative effects of flood disasters, researchers try to forecast flood by utilizing deep learning technology. Essentially, historical flood data can be regarded as sequential data with sets of flood factors. Facing challenges brought by inherent characteristics of flood forecasting, this paper proposes a dual attention embedding network, i.e., DA-Net, to achieve accurate prediction results. The proposed attention mechanism not only embeds a convolution self-attention module (CSA) on Temporal Convolutional Network (TCN) for description of local context information, but also constructs a Temporal-related Feature Attention (TFA) Module to assign time-varying weights for different features in a global sense. Specifically, CSA offers additional and local context information to help predict extreme runoff values even within a small period, meanwhile TFA improves global modeling capability of TCN for construction of data-driven generation mechanism in our method. Experiments on Changhua and Tunxi watershed dataset show the proposed method achieves superior prediction performance than current deep learning based methods.
2023
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11386/4825131
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 1
  • ???jsp.display-item.citation.isi??? 0
social impact