多要素局部-全局特征关联的有效波高预测模型

有效浪高(Significant Wave Heights，SWH)是描述海浪的重要属性，SWH预测对于保障近岸工程设计以及海上作业安全具有重要意义。近年来，深度学习方法被用来对SWH预测，但是目前存在的方法无法有效捕捉SWH的长时间相关性，同时忽略了海洋多要素之间的局部关联。为此，本文提出了一种结合海洋多要素局部和全局特征的SWH预测模型(Multi-elements Local and Global Correlation for Wave height Prediction, MLG-SWH)。首先，以有效浪高、风速、周期等多要素作为输入，设计了局部-全局编码(Local-Global Embedding，LGE)模块提取海洋多要素的局部关联以及时间信息；然后，采用编-解码器作为基础网络结构提取多要素海浪序列特征，在编-解码器中设计了因果空洞卷积自注意力模块有效捕捉海洋多要素序列的全局长时间相关性，并在解码器中利用生成推理方式避免单步迭代预测产生的误差累积；最后，选取北大西洋海浪浪高变化特点不同的两个站点数据进行实验。相较于经典时间序列预测模型以及主流深度学习方法，所提MLG-SWH模型在24、48小时预测的均方误差(MSE)以及平均绝对误差(MAE)均为得最低，并在长时序预测方面具有较大的优势。

A prediction model of significant wave height based on local and global correlation of multi-elements

Significant Wave Heights (SWH) is an important attribute to describe ocean waves, and SWH prediction is of great significance for ensuring the design of offshore engineering and the safety of offshore operations. In recent years, deep learning methods have been used to predict SWH, but the existing methods cannot effectively capture the long-term correlation of SWH, while ignoring the local associations between multiple elements of the ocean. To this end, this paper proposes a SWH prediction model (Multi-elements Local and Global Correlation for Wave height Prediction, MLG-SWH) that combines local and global features of marine multi-elements. First, using multiple factors such as significance wave height, wind speed as input, a Local-Global Embedding (LGE) module is designed to embed local correlation and periodic information of ocean multi-elements. Then, an encoder-decoder structure is used to extract the features of ocean wave height, where a casual dilated convolution and self-attention module is designed to effectively capture the global long-term correlation of ocean multi-element sequences and the generative reasoning prediction in the decoder is adopted to avoid errors accumulated in the single-step iterative prediction. Finally, the data of two stations with different characteristics of SWH variation in the North Atlantic are selected for experimental evaluations. Compared with classical time-series forecasting models and mainstream deep learning methods, the MLG-SWH model reaches the lowest mean square error (MSE) and mean absolute error (MAE) in 24 and 48 hours SWH forecasting, having a greater advantage in long-term time-series prediction.