基于Non stationary-CNN-Transformer的海浪有效波高预测

EFFECTIVE WAVE HEIGHT PREDICTION BASED ON NON STATIONARY-CNN-TRANSFORMER FOR OCEAN WAVES

针对海浪有效波高序列波动性、随机性较强,难以精确预测以及模型无法高效挖掘深层特征间关系的问题,提出一种基于Non stationary-CNN-Transformer模型的海浪有效波高预测方法。首先,使用平稳化模块减弱海浪时序数据的非平稳性;其次,利用一维卷积神经网络(CNN)提取相关数据间的深层特征并构建特征向量;最后,使用含有平稳性注意力的Transformer描述波高序列的时间依赖性捕捉到序列之间的全局关系,通过逆归一化处理后获得有效波高预测结果。该方法可消除海浪时序数据的非平稳性,提升数据的预测效果,并具有优异的特征提取能力且善于处理大规模时间序列数据。在实验中应用澳大利亚的浮标实测数据,通过7组对比实验分别预测0.5、3、6、12和24 h的有效波高,对所提模型进行全方位、多角度的验证。算例研究结果表明,该文所提模型在不同时间段精度有明显提升。

Aiming at the problems of wave effective height series volatility and strong randomness,which are difficult to predict accurately,and the model cannot efficiently explore the relationship between deep features,a method of wave effective height prediction based on non-stationary CNN-Transformer model is proposed.Firstly,a stabilization module is used to reduce the non-stationarity of wave time series data.Then,a one-dimensional convolutional neural network is used to extract deep features from the relevant data and create the feature vector.Finally,a Transformer with smoothing attention is used to depict the time-dependent wave height sequences and capture the overall relationship among the sequences.The effective prediction results of the wave height after the inverse normalization process are obtained.The approach eliminates the unstable nature of wave time series data,enhancing data prediction effects.Additionally,it possesses excellent feature extraction capabilities and can effectively handle large-scale time series data.During experiments,seven sets of comparative experiments were conducted with Australian buoys,predicting effective wave heights of 0.5,3,6,12 and 24 hours.The proposed model were thoroughly verified from all angles and in every aspect.The case study findings demonstrate that the precision of the model put forward in this paper is notably enhanced across various time periods.