Scholarlink

<p>Accurately predicting tidal levels in tidal reaches is crucial for the safety and sustainable development of estuarine regions.# Tidal levels in tidal reaches are influenced by natural factors, including upstream runoff and ocean tides, as well as anthropogenic activities, making their prediction particularly challenging. In recent years, deep neural networks have shown great potential in tidal level predictions due to their strong nonlinear mapping capabilities, generalization abilities, and adaptability. This study employed one convolutional neural network (CNN) and three different recurrent neural networks (RNNs) to construct various tidal level prediction models, including four single-structure models, three serial CNN–RNN models, and three parallel-serial hybrid CNN–RNN models (iCNN–RNN).# These ten models were used to predict tidal levels at Wenzhou Station in the tidal reach of the Oujiang River Basin in East China under eight different input conditions. Model performance was comprehensively evaluated with a focus on both the overall accuracy and the ability to predict extreme high tidal levels. Results showed that the single-structure RNN models generally outperformed the CNN models in tidal level predictions. The serial CNN–RNN models did not show significant advantages over single-structure models in predicting tidal levels. However, the parallel-serial hybrid iCNN–RNN models demonstrated superior overall performance across all tidal levels with higher <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(NSE\)</EquationSource> </InlineEquation> values and lower <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(RMSE\)</EquationSource> </InlineEquation> values compared to other models; moreover, they also exceled in extreme high tidal level predictions based on the <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\({RMSE}_{H}\)</EquationSource> </InlineEquation>, POD, FAR, and CSI metrics. The superiority of the iCNN–RNN models for tidal level predictions result from their advantageous structures, which enabled the CNN and RNN to simultaneously extract features and temporal correlations from the original input data, optimally exploiting the strengths of both the CNN and the RNN.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid iCNN–RNN model for enhanced tidal level prediction: comparative analysis of multiple deep neural networks

  • 期刊论文

  • 发表日期：

    2025年9月17日

查看全文
  • Zhixu Bai,
  • Yutai Ke,
  • Di Ma,
  • Qianwen Wu,
  • Suli Pan

摘要

Accurately predicting tidal levels in tidal reaches is crucial for the safety and sustainable development of estuarine regions.# Tidal levels in tidal reaches are influenced by natural factors, including upstream runoff and ocean tides, as well as anthropogenic activities, making their prediction particularly challenging. In recent years, deep neural networks have shown great potential in tidal level predictions due to their strong nonlinear mapping capabilities, generalization abilities, and adaptability. This study employed one convolutional neural network (CNN) and three different recurrent neural networks (RNNs) to construct various tidal level prediction models, including four single-structure models, three serial CNN–RNN models, and three parallel-serial hybrid CNN–RNN models (iCNN–RNN).# These ten models were used to predict tidal levels at Wenzhou Station in the tidal reach of the Oujiang River Basin in East China under eight different input conditions. Model performance was comprehensively evaluated with a focus on both the overall accuracy and the ability to predict extreme high tidal levels. Results showed that the single-structure RNN models generally outperformed the CNN models in tidal level predictions. The serial CNN–RNN models did not show significant advantages over single-structure models in predicting tidal levels. However, the parallel-serial hybrid iCNN–RNN models demonstrated superior overall performance across all tidal levels with higher \(NSE\) values and lower \(RMSE\) values compared to other models; moreover, they also exceled in extreme high tidal level predictions based on the \({RMSE}_{H}\) , POD, FAR, and CSI metrics. The superiority of the iCNN–RNN models for tidal level predictions result from their advantageous structures, which enabled the CNN and RNN to simultaneously extract features and temporal correlations from the original input data, optimally exploiting the strengths of both the CNN and the RNN.