Objective Accurate prediction of water depth in coastal waterways is essential for ensuring the safety and efficiency of construction and transportation activities, particularly in environments characterized by complex and dynamic hydrological conditions. The Pinglu Canal, an inland waterway affected by tidal influences, serves as an example where traditional prediction models fail to accurately forecast water depth under complex hydrological conditions. This study presents a water depth prediction model based on an attention-enhanced long short-term memory (LSTM) network, which is integrated into a decision-support platform for real-time channel transportation management. Methods First, key hydrological factors, including upstream discharge, daily rainfall, tidal current velocity, and tidal level, were incorporated to construct the LSTM-based coastal waterway depth prediction model. The raw hydrological data were preprocessed to address issues such as missing values, noise, and irregular time intervals, ensuring suitability for time-series modeling. Then, the LSTM model was utilized to capture the long-term temporal dependencies within the data, allowing the model to account for complex interactions among various hydrological variables over time. An attention mechanism was introduced to optimize the model's performance. This mechanism improved the model architecture by enabling it to dynamically adjust the weight of each feature at every time step, prioritizing the most relevant factors based on the current data. The attention mechanism enhanced both the accuracy and stability of the model, particularly for ultra-long-term water depth forecasting under dynamic and complex coastal hydrological conditions. Finally, the optimized model was embedded into a transportation decision-support platform, allowing for real-time water depth prediction, dynamic correction of forecasts based on new data, and navigable time window evaluation. The model's effectiveness was validated through comparative analyses with existing prediction models and field measurements, demonstrating superior accuracy and reliability in predicting waterway depth. Results and Discussions The results demonstrated that at two monitoring points, 5 km from the coast (Point 1) and 30 km inland (Point 2), the traditional LSTM model exhibited larger prediction errors, especially in long-term forecasts. The MAE ranged from 0.07 m to 1.08 m for short-term predictions and from 0.12 m to 1.74 m for long-term forecasts. The model also tended to overestimate water depth. In contrast, the attention mechanism-based LSTM model consistently kept the MAE below 0.15 m, even under sudden rainfall or upstream discharge events, showing enhanced accuracy in both short-term fluctuations and long-term trends. The model's performance across seasonal variations further highlighted its robustness. During the dry season, the MAE was reduced by 64.68%, and during the wet season, it decreased by 72.36%. The RMSE was also reduced by 67.51% and 73.39% in the respective seasons, while the R² coefficient improved by 2.18% and 5.60%. This demonstrated the model's adaptability to both stable and volatile water conditions. The attention mechanism-based LSTM model significantly outperformed traditional LSTM models in predicting waterway depth. Compared to the traditional model, the MAE of was reduced by 65.00%~72.00%, and the R² coefficient increased by 2.20%~5.60%, demonstrating superior predictive accuracy and stability. This improvement was particularly evident under complex hydrological conditions, where the model effectively captured non-linear and dynamic relationships among key features such as tidal flow speed, daily rainfall, and tidal water level. In addition, when compared to a single feature vector model, the three-feature vector combination (daily rainfall, tidal flow speed, and tidal water level) resulted in an MAE of no more than 0.15 m and an R² coefficient of no less than 0.99, substantially improving the model's accuracy and stability for predicting waterway depth under complex coastal hydrological conditions. Finally, when integrated into the waterway transportation decision-support platform, the model's capabilities, such as real-time water depth prediction, dynamic correction, and navigable time window evaluation, substantially enhanced the platform's effectiveness. This integrated system provided reliable and accurate information for waterway transportation management, contributing to safer and more efficient navigation in complex coastal environments. Conclusions This study presents an attention mechanism-based LSTM model for predicting waterway depth in complex coastal environments. The proposed model substantially improves prediction accuracy and stability, especially under dynamic hydrological conditions. The model effectively adapts to diverse waterway conditions by incorporating key hydrological features, ensuring higher precision in depth forecasting. When implemented within a waterway transportation decision-support platform, the model enables real-time prediction, dynamic correction, and navigable time window assessment, enhancing the intelligence and digital management of waterway transportation systems. This research provides reliable technical support for future engineering applications.
为进一步分析输入特征数量对模型预测精度的影响,本文通过不同数量的特征输入进行分析,分别使用包含上游流量、日降雨量、潮流流速和潮汐水位4个特征向量(分别简记为 S 、 R 、 C 、 X )及待测点水深(W)为标签进行训练并评估模型精度。数据集仍以第3.1节的样本为例。通过逐步增加输入特征数量,本文采用不同的组合方式,分别按特征数量1~4进行组合,共得到15种组合方式。图8和图9分别为EMAE和R2雷达图,采用极坐标系,直径越大,EMAE和R2也越大。由图8可见,随着特征向量数量的增加,MAE的值呈现逐渐减小的趋势。其中,由图8(a)可见,在枯水期时, S ‒ R ‒ C ‒ X 、 R ‒ C ‒ X 、 S ‒ C ‒ X 和 C ‒ X 组合下EMAE均较小,分别为0.07、0.14、0.15和0.17。反之,其他特征组合的EMAE普遍较大,且均不小于0.66,表明在枯水期水文地质环境下潮流流速和潮汐水位对预测精度的影响较为显著,而上游流量和日降雨量的影响权重较小。由图8(b)可见,在丰水期时, S ‒ R ‒ C ‒ X 、 S ‒ R ‒ X 、 S ‒ C ‒ X 和 R ‒ C ‒ X 组合的MAE值均较小,分别0.04、0.10、0.17和0.15。相比之下,其他特征组合的EMAE均较大,且整体均不小于0.33,表明丰水期水文地质环境条下潮汐水位对模型预测精度的影响较为显著,且上游流量和日降雨量二者影响相同。由图9可见,在单一特征输入时,模型的预测精度较低,决定系数R²约为0.50。当增加第3个特征时,模型精度明显提高,R²均超过0.91,且当特征向量按日降雨量/上游流量、潮流流速和潮汐水位特征组合时,R²均不低于0.99。然而,当加入第4个特征(即日降雨量或上游流量)后,R²稳定在0.995,说明加入日降雨量或上游流量作为第4个输入特征对模型精度的提升不显著。
XingChangli.Discussion on supervision mode of water traffic safety of Hong Kong‒Zhuhai‒Macao Bridge[J].Railway Engineering,2014,54(7):15‒18.
[2]
邢长利.港珠澳大桥水上交通安全监管模式探讨[J].铁道建筑,2014,54(7):15‒18.
[3]
LiuHuaihan, YangShengfa, CaoMinxiong.Advances in 'golden waterway' regulation technologies of the Yangtze River[J].Advanced Engineering Sciences,2017,49(2):17‒27.
XiaolongLyu, WuPeng, LiuXiaoling.Demonstration on waterway classification of Pinglu canal[J].Port & Waterway Engineering,2021(10):266‒270.
[6]
吕小龙,吴澎,刘晓玲.平陆运河航道等级论证[J].水运工程,2021(10):266‒270.
[7]
GanMin, ChenYongping, PanShunqi,et al.A modified nonstationary tidal harmonic analysis model for the Yangtze estuarine tides[J].Journal of Atmospheric and Oceanic Technology,2019,36(4):513‒525. doi:10.1175/jtech-d-18-0199.1
[8]
BarzegarR, FijaniE, Asghari MoghaddamA,et al.Forecasting of groundwater level fluctuations using ensemble hybrid multi-wavelet neural network-based models[J].Science of the Total Environment,2017,599:20‒31. doi:10.1016/j.scitotenv.2017.04.189
[9]
SongyiDian, LiXiaoying, YangDan,et al.Adaptive robust prediction of groundwater level based on fusion attention mechanism LSTM network[J].Advanced Engineering Sciences,2024,56(1):54‒64.
GribovszkiZ, KaliczP, SzilágyiJ,et al.Riparian zone evapotranspiration estimation from diurnal groundwater level fluctuations[J].Journal of Hydrology,2008,349(1/2):6‒17. doi:10.1016/j.jhydrol.2007.10.049
[14]
ChenYifan, CaoFeifeng, MengXiangyong,et al.Water level simulation in river network by data assimilation using ensemble Kalman filter[J].Applied Sciences,2023,13(5):3043. doi:10.3390/app13053043
[15]
AlmikaeelW, ŠoltészA, ČubanováL,et al.Hydro-informer:A deep learning model for accurate water level and flood predictions[J].Natural Hazards,2025,121(4):3959‒3979. doi:10.1007/s11069-024-06949-8
[16]
MaXin, HuHao, RenYufeng.A hybrid deep learning model based on feature capture of water level influencing factors and prediction error correction for water level prediction of cascade hydropower stations under multiple time scales[J].Journal of Hydrology,2023,617:129044. doi:10.1016/j.jhydrol.2022.129044
[17]
ZhangQixiao, LiPeiyue, RenXiaofei,et al.A new real-time groundwater level forecasting strategy:Coupling hybrid data-driven models with remote sensing data[J].Journal of Hydrology,2023,625:129962. doi:10.1016/j.jhydrol.2023.129962
[18]
SunWei, ChangL C, ChangF J.Deep dive into predictive excellence:Transformer's impact on groundwater level prediction[J].Journal of Hydrology,2024,636:131250. doi:10.1016/j.jhydrol.2024.131250
[19]
HaifengLyu, JiXiaoyu, DingYong.Research on Xijiang River water level prediction model of multi-factor based on splice-LSTM[J].Yangtze River,2023,54(7):81‒88.
LiuXiaolei, LiuYifan, WangShuaifei,et al.New canal construction and marine emissions strategy:A case of Pinglu[J].Frontiers in Marine Science,2024,10:1303879. doi:10.3389/fmars.2023.1303879
[22]
XinJingzhou, ZhouChaoying, JiangYan,et al.A signal recovery method for bridge monitoring system using TVFEMD and encoder-decoder aided LSTM[J].Measurement,2023,214:112797. doi:10.1016/j.measurement.2023.112797
[23]
Martin NascimentoG F, WurtzF, Kuo‒PengP,et al.Outlier detection in buildings' power consumption data using forecast error[J].Energies,2021,14(24):8325. doi:10.3390/en14248325
[24]
MaRuixin, YinYong, BaoKexin,et al.Water level prediction of inland waterways based on MHA-BiGRU[J].Journal of Dalian Maritime University,2024,50(1):46‒56.
TianZhongda, YuXiyan, FengGuokui.Short-term wind speed prediction model based on long short-term memory network with feature extraction[J].Earth Science Informatics,2025,18(4):333. doi:10.1007/s12145-025-01851-8
[27]
ZhangQi, MiaoChiyuan, GouJiaojiao,et al.Spatiotemporal characteristics and forecasting of short-term meteorological drought in China[J].Journal of Hydrology,2023,624:129924. doi:10.1016/j.jhydrol.2023.129924
[28]
HeGuangxin, QuHaifeng, LuoJingjia,et al.An long short-term memory model with multi-scale context fusion and attention for radar echo extrapolation[J].Remote Sensing,2024,16(2):376. doi:10.3390/rs16020376
[29]
LiHaoran, ZhangLili, ZhangYaowen,et al.Water-level prediction analysis for the Three Gorges Reservoir area based on a hybrid model of LSTM and its variants[J].Water,2024,16(9):1227. doi:10.3390/w16091227
HuangYing, XuJian, ZhouZiqi,et al.Research and implementation of efficient long sequence model for water level forecasting[J].Journal of University of Electronic Science and Technology of China,2023,52(4):595‒601.
BahdanauD, ChoK, BengioY.Neural machine translation by jointly learning to align and translate[EB/OL].(2015‒04‒07)[2025‒01‒15].doi:10.1017/9781108608480.003
[37]
ZengYan, FengDongming, LiJian'an.Vehicle-assisted bridge damage assessment by combining attention mechanism and Bi-LSTM network[J].Journal of Vibration Engineering,2024,37(7):1089‒1097.