本文以 2019年 1月 1日至 2021年 12月 31日舟山群岛南部外海观测点所涵盖的气象、海洋、地形等多种物理量数据为数据基础,使用长短时记忆(Long Short Term Memory,LSTM)神经网络搭建深度学习海浪预报模型,探讨输入输出序列比和输入要素数量对模型预测性能的影响,在舟山海域实现波浪三要素,即有效波高、有效波周期、传播方向的短时预报,并用 2022年台风"轩岚诺"和"梅花"期间的数据检验模型对极端海况的预测能力.研究结果表明,根据实测数据所训练的多要素海浪预报模型具有较好的预测准确度和稳定性,能较好地实现对极端海况的预测,当输入输出序列比为1∶1时模型准确度较高,预报时长为 1 h的三要素模型对于日常海况中有效波高、有效波周期和波向的预测均方根误差(Root Mean Squared Error,RMSE)分别为 0.116 m、0.569 s和 24.583°,对于极端海况中有效波高的预测RMSE为 0.191 m,输入要素数量的增加可进一步提升模型准确度,但在预测时长较长时也会增加训练成本.
A multivariate wave forecasting model for the Zhoushan archipelago using Long Short-Term Memory deep neural networks
This study is based on the meteorological,oceanic,terrain and other physical quantity data covered by the observation points in the southern Zhoushan Islands from January 1,2019 to December 31,2021,and uses long short-term memory neural network(LSTM)to build deep learning wave forecast model.We explore the impact of the input-output sequence ratio and the number of input elements on the prediction performance of the model,real-ize the short-term forecast of the three elements of waves in the Zhoushan sea area,that is the significant wave height,the significant wave period and the propagation direction,and use the data during the 2022 typhoons"Hin-namnor"and"Muifa"to test the model's prediction ability for extreme sea conditions.The research results show that the multi-element deep learning wave forecast model trained based on measured data has good prediction accur-acy and stability,and can realize the prediction of extreme sea conditions.When the input-output sequence ratio is 1∶1,the model accuracy is higher.In non-extreme sea conditions,the three-element model with a prediction time of 1 hour accurately predicts significant wave height,significant wave period and direction,with Root Mean Squared Errors(RMSE)of 0.116 m,0.569 s,and 24.583° respectively.In extreme sea conditions,the prediction RMSE for the significant wave height is 0.191 m.The increase in the number of input elements can further improve the model accuracy but also increase the training cost when the prediction time is long.
deep learningLong Short-Term Memory modelwave forecastingZhoushan
NETL
NSTL
万方数据
