桩柱是浅海和近岸工程结构的重要支撑构件。桩基周围海床在海浪或水流作用下的冲刷深度是一个重要的结构稳定设计参数，对其准确预测具有重要的工程意义和经济价值。目前，局部冲刷深度预测普遍采用经验公式、数学模型及人工智能方法。经验公式法包含的影响因素不完全，适用范围有限；而数学模型往往需要依赖确定复杂的动力地貌演变过程，计算量大，不便于工程设计使用。近年来，各种人工智能算法，特别是人工神经网络（artificial neural network，简称 ANN）方法，已经被应用到桩基周围局部冲刷深度计算，显示出了优越的预测能力。应用多层感知机反向传播算法神经网络方法（MLP/BP）建立了预测波、流分别作用下桩基局部平衡冲刷深度模型。模型比较了采用有量纲和无量纲训练参数数据输入得到的预测精度，并通过系统的敏感性分析，确定了波流参数和泥沙特征对计算结果的影响程度。研究结果不仅证实了无论是对应波浪还是水流作用条件，神经网络模型均优于大多数现有工程使用的经验公式，还证实了采用有量纲参数输入训练的模型可以得到比无量纲输入模型更为准确的预测结果。

英文摘要:

Piles are essential supporting components of shallow-sea and nearshore engineering structures. The local scour depth around pile foundations under wave or current action is a critical design parameter for structural stability, and its accurate prediction has significant engineering and economic value. Currently, the commonly adopted methods for predicting local scour depth include empirical formulas, mathematical models, and artificial intelligence (AI) models. Empirical formulas are usually limited by incomplete consideration of influencing factors and narrow applicability, while mathematical models often rely on complex morphodynamic evolution, which requires substantial computational resources and is thus less practical for engineering design. Recently, various AI methods, especially artificial neural networks (ANN), have been applied to the prediction of local scour depth around pile foundations, demonstrating superior predictive performance. This study employs a multilayer perceptron back-propagation (MLP/BP) algorithm neural network to establish models for predicting local equilibrium scour depth around pile foundations under both wave and current actions. Comparisons of prediction accuracy are conducted using dimensional and non-dimensional parameters as training data. The influences of wave and current conditions as well as sediment characteristics are systematically assessed through sensitivity analyses. The results confirm that for both wave and current cases, the neural network model outperforms most empirical formulas widely used in engineering practice. Moreover, the ANN model trained with dimensional parameter inputs achieves more accurate predictions than those trained with non-dimensional inputs.

查看全文查看/发表评论下载PDF阅读器

关闭