A Physically Interpretable Dual‐Encoder Cross‐Attention LSTM for Water Level Forecasting in Dam–Tide Interaction Systems

ABSTRACT Predicting the water level in real time in regulated tidal rivers is challenging because upstream dam releases and downstream tidal fluctuations act simultaneously, with their relative dominance shifting within an event. Attention‐based deep learning models for such systems have generally either mixed all forcings within a single encoder, meaning no forcing‐level attribution can be recovered, or distributed attention across variable–lag pairs in a self‐attention stream. This makes it impossible to read the relative influence of two named physical forcings. This study proposes a dual‐encoder cross‐attention long short‐term memory (LSTM) framework that separates the forcings at the encoder level rather than the attention level. Dam discharge and tidal stage are encoded by dedicated bidirectional long short‐term memory (BiLSTM) encoders. Attention is then applied at two levels: temporal cross‐attention with a per‐horizon query and channel attention normalized over the two forcings. When applied to the Gahwa River Basin, which is subject to releases from the Namgang Dam and tidal forcing from Sacheon Bay, the framework was evaluated across 12 flood events using event‐based cross‐validation. Predictions from 1 to 12 h ahead were benchmarked against those from a single‐encoder LSTM. The two models performed comparably at short lead times. Beyond 5–6 h, however, the baseline deteriorated and became unstable. For Event #2, the Nash–Sutcliffe efficiency (NSE) declined to 0.279 at 12 h, and for Event #9 it fluctuated between 0.33 and 0.57. In contrast, the proposed model retained values of 0.425 and 0.627, respectively, and degraded monotonically. The normalized channel weights provide a clear indication of the influence at each time step, and their semidiurnal oscillation in phase with the tidal record suggests that the model is organized around the correct physical periodicity.

Authors

Institutions

Publication Details

Journal
Journal of Flood Risk Management
Published
2026-11-15
DOI
https://doi.org/10.1111/jfr3.70284
Primary Topic
Hydrological Forecasting Using AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Physically Interpretable Dual‐Encoder Cross‐Attention LSTM for Water Level Forecasting in Dam–Tide Interaction Systems

Boosik Kang, Jiyoung Sung
Journal of Flood Risk Management
Hydrological Forecasting Using AI
article

A Physically Interpretable Dual‐Encoder Cross‐Attention LSTM for Water Level Forecasting in Dam–Tide Interaction Systems

Boosik Kang, Jiyoung Sung
article en

Abstract

ABSTRACT Predicting the water level in real time in regulated tidal rivers is challenging because upstream dam releases and downstream tidal fluctuations act simultaneously, with their relative dominance shifting within an event. Attention‐based deep learning models for such systems have generally either mixed all forcings within a single encoder, meaning no forcing‐level attribution can be recovered, or distributed attention across variable–lag pairs in a self‐attention stream. This makes it impossible to read the relative influence of two named physical forcings. This study proposes a dual‐encoder cross‐attention long short‐term memory (LSTM) framework that separates the forcings at the encoder level rather than the attention level. Dam discharge and tidal stage are encoded by dedicated bidirectional long short‐term memory (BiLSTM) encoders. Attention is then applied at two levels: temporal cross‐attention with a per‐horizon query and channel attention normalized over the two forcings. When applied to the Gahwa River Basin, which is subject to releases from the Namgang Dam and tidal forcing from Sacheon Bay, the framework was evaluated across 12 flood events using event‐based cross‐validation. Predictions from 1 to 12 h ahead were benchmarked against those from a single‐encoder LSTM. The two models performed comparably at short lead times. Beyond 5–6 h, however, the baseline deteriorated and became unstable. For Event #2, the Nash–Sutcliffe efficiency (NSE) declined to 0.279 at 12 h, and for Event #9 it fluctuated between 0.33 and 0.57. In contrast, the proposed model retained values of 0.425 and 0.627, respectively, and degraded monotonically. The normalized channel weights provide a clear indication of the influence at each time step, and their semidiurnal oscillation in phase with the tidal record suggests that the model is organized around the correct physical periodicity.

Journal of Flood Risk ManagementVol. 19(4)
Dankook University (KR)
Hydrological Forecasting Using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon