南海岛礁海域波浪智能预报方法应用研究

王发云; 王盛健; 陈斌; 梁乃安; 邱立国

doi:10.12284/hyxb20250111

南海岛礁海域波浪智能预报方法应用研究

doi: 10.12284/hyxb20250111

海南省海洋与渔业科学院，海南海口 571126

基金项目: 海南省海洋与渔业科学院科研项目结余经费课题（JYJF20231226）。

详细信息

作者简介:
王发云（1986—），男，海南省屯昌县人，高级工程师，主要从事海洋动力过程研究。E-mail：95309728@qq.com

通讯作者:
邱立国，男，副研究员，主要从事海洋动力过程研究。E-mail：389648810@qq.com

中图分类号: P731.33
计量
- 文章访问数: 121
- HTML全文浏览量: 62
- PDF下载量: 13
- 被引次数: 0
出版历程
- 收稿日期: 2025-08-25
- 修回日期: 2025-11-13
- 网络出版日期: 2025-11-22
- 刊出日期: 2025-12-31

Application of an intelligent wave forecasting method to waters around the islands and reefs in the South China Sea

Hainan Academy of Ocean and Fisheries Sciences, Haikou 571126, China

摘要

摘要: 为提升南海岛礁海域的波浪预报精度与模型泛化能力，本文基于BO-LSTM模型，系统探讨了输入因素作用、模型跨站位迁移能力及多步预测方法性能。研究采用单因素（历史波高）与双因素（历史波高+风速）输入，通过优化时间窗口，结合滚动预测法（RF）与直接多步预测法（DM），对七连屿、甘泉岛、晋卿岛和华夏暗沙4个站位的1～24 h有效波高进行预报验证。结果表明：模型表现与站点地理地貌揭示的水动力环境高度相关。七连屿站“半遮蔽−半开阔”的格局使其模型泛化能力最强且稳定（最佳窗口n = 2），在跨站预测中表现优异；而甘泉岛站、晋卿岛站等“潟湖内局地型”站点与华夏暗沙站“开阔水域型”站点，则因其地理特征主导的、差异化的数据分布导致了显著的“数据域偏移”，限制了模型的跨站迁移能力。短期预报中历史波高为核心因素，但其优势存在地理依赖性：七连屿站与晋卿岛站历史波高权重优势显著（>1.7倍），而甘泉岛站与华夏暗沙站风速贡献相对提升（优势比<1.4倍）。多步预测中，“DM+双”在短期至中期（1～18 h）综合最优，“RF+双”在长期（19～24 h）及甘泉岛站全时段更能抑制误差衰减。本研究验证了BO-LSTM在南海岛礁波浪预报中的有效性，并通过关联数据驱动规律与地理物理机制，为构建区域普适性智能预报模型提供了物理可解释的见解与方法支撑。
- 波浪预报 /
- BO-LSTM /
- 南海岛礁海域 /
- 跨站位迁移 /
- 数据域偏移
Abstract: This study aims to enhance wave forecast accuracy and model generalization in the South China Sea island and reef waters using a BO-LSTM model. We systematically investigated the effects of input factors, model cross-station transferability, and multi-step prediction performance. The research employed single-factor (historical wave height) and dual-factor (historical wave height + wind speed) input schemes. Forecasts for 1−24 h were generated and validated at four stations (Qilianyu, Ganquan Island, Jinqing Island, and Huaxia Shoal) using the Rolling Forecast (RF) and Direct Multi-step (DM) methods. Results show that model performance is highly correlated with the hydrodynamic environment dictated by station geomorphology. The model trained on data from Qilianyu Station, with its “semi-sheltered to semi-open” setting, demonstrated the strongest and most stable generalization ability (optimal window n = 2) and excellent cross-station performance. In contrast, models for the “localized lagoon” stations (Ganquan Island, Jinqing Island) and the “open water” station (Huaxia Shoal) exhibited limited transferability due to significant “data domain shift” arising from their distinct geographic settings. For short-term forecasts, historical wave height was the core input factor, but its dominance showed geographic dependence. Its weight was significantly higher (>1.7 times) than wind speed at Qilianyu and Jinqing Island, while wind speed contribution was greater (advantage ratio <1.4) at Ganquan Island and Huaxia Shoal. For multi-step forecasting, “DM + Dual” performed best for short-to-medium terms (1−18 h), whereas “RF + Dual” was superior for long-term forecasts (19−24 h) and at Ganquan Island across all horizons. This study validates BO-LSTM’s effectiveness for wave forecasting in the South China Sea and provides physically interpretable insights for developing regional intelligent forecasting models by linking data-driven patterns with geophysical mechanisms.
- wave forecasting /
- BO-LSTM /
- island and reef waters in the South China Sea /
- cross-station transfer /
- data domain shift

HTML全文

图 1 研究区域及其波浪观测站位置示意图

Fig. 1 The study area and the indication of the wave measurement station

下载: 全尺寸图片幻灯片

图 2 数据预处理流程

Fig. 2 The process diagram of data preprocessing

下载: 全尺寸图片幻灯片

图 3 各个波浪观测站有效波高预测结果

a. 七连屿站; b. 甘泉岛站; c. 晋卿岛站; d. 华夏暗沙站

Fig. 3 Prediction results of the significant wave height at various wave measurement stations

a. Qilianyu Station; b. Ganquandao Station; c. Jinqingdao Station; d. Huaxiaansha Station

下载: 全尺寸图片幻灯片

图 4 LSTM单元结构

Fig. 4 LSTM unit structure

下载: 全尺寸图片幻灯片

图 5 岛礁海域波浪预报模型流程

Fig. 5 The construction process diagram of wave forecasting model for island and reef area

下载: 全尺寸图片幻灯片

图 6 两种输入因素不同时间窗口下各组次的RMSE

a. 输入单因素；b. 输入双因素

Fig. 6 RMSE of each group under different time windows for two input factors

a. Single-factor input; b. two-factor input

下载: 全尺寸图片幻灯片

图 7 本站位模型预测精度时序对比

a. 七连屿站；b. 甘泉岛站；c. 晋卿岛站；d. 华夏暗沙站

Fig. 7 Time-series validation of model prediction accuracy at the current station

a. Qilianyu Station; b. Ganquandao Station; c. Jinqingdao Station; d. Huaxiaansha Station

下载: 全尺寸图片幻灯片

图 8 有效波高核密度散点图

a. 七连屿站; b. 甘泉岛站; c. 晋卿岛站; d. 华夏暗沙站

Fig. 8 Kernel density scatter plots of significant wave height

a. Qilianyu Station; b. Ganquandao Station; c. Jinqingdao Station; d. Huaxiaansha Station

下载: 全尺寸图片幻灯片

图 9 不同预测法评价指标对比时序对比图

a. 七连屿站; b. 甘泉岛站; c. 晋卿岛站; d. 华夏暗沙站

Fig. 9 Comparison of evaluation indicators for different forecasting methods using time series comparison chart

a. Qilianyu Station; b. Ganquandao Station; c. Jinqingdao Station; d. Huaxiaansha Station

下载: 全尺寸图片幻灯片

图 10 不同数据集输入因素权重绝对值均值的小提琴图

Fig. 10 Violin plot of mean values of absolute weights of input factors across different datasets

下载: 全尺寸图片幻灯片

表 1 波浪观测站信息^[22]

Tab. 1 Wave measurement station information

站位	观测仪器	观测时间（年/月）	采样间隔/h	样本量
七连屿站	ADCP	2017/07−2017/11	2	1371
甘泉岛站	浪龙仪	2018/01−2018/04	1	2603
晋卿岛站	浪龙仪	2014/06−2018/06	0.5	32487
华夏暗沙站	ADCP	2018/06−2018/11	2	1506

下载: 导出CSV

表 2 各模型预测误差对比

Tab. 2 Error comparison of prediction models

模型	七连屿站		甘泉岛站		晋卿岛站		华夏暗沙站
模型	RMSE/m	PCC	RMSE/m	PCC	RMSE/m	PCC	RMSE/m	PCC
LR	0.189	0.875	0.115	0.929	0.103	0.885	0.297	0.923
RFR	0.061	0.988	0.030	0.996	0.034	0.989	0.103	0.991
LSTM	0.149	0.927	0.098	0.949	0.098	0.898	0.244	0.949

下载: 导出CSV

表 3 最佳时间窗口

Tab. 3 Best time window

站位	测试集 RMSE/m	测试集 PCC	最佳时间窗口n	输入因数
七连屿站	0.036	0.958	2	历史波高
	0.037	0.956	2	历史波高+风速
甘泉岛站	0.042	0.989	1	历史波高
	0.042	0.990	1	历史波高+风速
晋卿岛站	0.047	0.977	1	历史波高
	0.048	0.976	1	历史波高+风速
华夏暗沙站	0.147	0.988	2	历史波高
	0.157	0.986	1	历史波高+风速

下载: 导出CSV

表 4 两种输入因素不同站位模型超参数

Tab. 4 Two input factors with different station model hyperparameters

站位	层数	学习率	隐藏层	迭代次数	输入因数
七连屿站	4	0.001	80	143	历史波高
七连屿站	4	0.003	93	132	历史波高+风速
甘泉岛站	3	0.041	73	193	历史波高
甘泉岛站	2	0.063	75	112	历史波高+风速
晋卿岛站	3	0.041	73	193	历史波高
晋卿岛站	3	0.041	73	193	历史波高+风速
华夏暗沙站	3	0.041	73	193	历史波高
华夏暗沙站	3	0.041	73	193	历史波高+风速

下载: 导出CSV

表 5 最优模型预测结果统计

Tab. 5 Statistics of optimal model prediction results

最优模型	输入因素	七连屿		甘泉岛		晋卿岛		华夏暗沙
最优模型	输入因素	RMSE/m	PCC	RMSE/m	PCC	RMSE/m	PCC	RMSE/m	PCC
七连屿数据（训练）	单因素	0.066	0.968	0.059	0.978	0.056	0.964	0.177	0.979
七连屿数据（训练）	双因素	0.066	0.969	0.059	0.980	0.055	0.966	0.183	0.979
甘泉岛数据（训练）	单因素	0.090	0.941	0.046	0.987	0.047	0.975	0.560	0.885
甘泉岛数据（训练）	双因素	0.090	0.944	0.044	0.988	0.051	0.974	0.531	0.908
晋卿岛数据（训练）	单因素	0.080	0.953	0.046	0.987	0.046	0.976	0.395	0.950
晋卿岛数据（训练）	双因素	0.085	0.948	0.046	0.988	0.046	0.975	0.492	0.926
华夏暗沙数据（训练）	单因素	0.143	0.94	0.102	0.978	0.158	0.931	0.143	0.986
华夏暗沙数据（训练）	双因素	0.167	0.949	0.121	0.986	0.158	0.969	0.151	0.984

下载: 导出CSV

表 6 不同预测法评价指标对比

Tab. 6 Comparison of evaluation metrics among different prediction methods

模型	预报时序/h	七连屿站			甘泉岛站			晋卿岛站			华夏暗沙站
模型	预报时序/h	NMAE	NRMSE	PCC	NMAE	NRMSE	PCC	NMAE	NRMSE	PCC	NMAE	NRMSE	PCC
RF+单	1	0.009	0.015	0.968	0.015	0.026	0.987	0.014	0.019	0.975	0.016	0.025	0.984
	3	0.012	0.021	0.938	0.028	0.044	0.961	0.018	0.026	0.938	0.025	0.035	0.968
	6	0.015	0.027	0.895	0.04	0.062	0.922	0.024	0.035	0.89	0.033	0.046	0.943
	12	0.021	0.038	0.779	0.056	0.086	0.846	0.031	0.044	0.811	0.046	0.064	0.889
	24	0.029	0.054	0.527	0.076	0.116	0.713	0.04	0.057	0.667	0.066	0.091	0.767
RF+双	1	0.009	0.015	0.969	0.016	0.025	0.988	0.011	0.016	0.977	0.016	0.025	0.984
	3	0.012	0.021	0.940	0.028	0.041	0.968	0.018	0.026	0.937	0.026	0.035	0.967
	6	0.016	0.027	0.900	0.038	0.054	0.946	0.025	0.035	0.888	0.035	0.046	0.945
	12	0.023	0.037	0.795	0.049	0.068	0.919	0.031	0.044	0.818	0.045	0.059	0.906
	24	0.032	0.054	0.555	0.061	0.083	0.889	0.038	0.054	0.736	0.055	0.071	0.860
DM+单	1	0.009	0.015	0.968	0.015	0.026	0.987	0.014	0.019	0.975	0.016	0.025	0.984
	3	0.013	0.022	0.929	0.028	0.044	0.961	0.018	0.026	0.939	0.025	0.036	0.966
	6	0.015	0.027	0.897	0.039	0.061	0.923	0.023	0.034	0.891	0.033	0.046	0.943
	12	0.019	0.035	0.813	0.054	0.084	0.852	0.029	0.043	0.823	0.046	0.063	0.891
	24	0.027	0.049	0.597	0.074	0.110	0.728	0.037	0.054	0.704	0.064	0.087	0.776
DM+双	1	0.009	0.015	0.969	0.016	0.025	0.988	0.011	0.016	0.977	0.016	0.025	0.984
	3	0.014	0.022	0.935	0.027	0.04	0.969	0.018	0.026	0.938	0.024	0.033	0.971
	6	0.017	0.026	0.905	0.035	0.052	0.945	0.022	0.032	0.905	0.030	0.040	0.957
	12	0.023	0.037	0.8	0.047	0.07	0.898	0.028	0.041	0.842	0.039	0.052	0.926
	24	0.030	0.048	0.614	0.065	0.098	0.788	0.036	0.053	0.718	0.054	0.073	0.847

下载: 导出CSV

表 7 不同数据集输入因素的模型权重强度对比

Tab. 7 Comparison of model weight intensity of input factors across different datasets

数据集	风速		历史波高		历史波高相对风速的优势（历史波高均值/风速均值）
数据集	平均权重强度	模型间标准差	平均权重强度	模型间标准差	历史波高相对风速的优势（历史波高均值/风速均值）
Q站	0.197	0.078	0.339	0.125	1.72
G站	0.188	0.065	0.216	0.103	1.14
J站	0.149	0.052	0.323	0.111	2.17
H站	0.193	0.059	0.262	0.125	1.36
注：平均权重强度指1～24 h预报模型的权重绝对值均值（单模型内先平均，再取24个模型的均值）；模型间标准差反映24个模型权重强度的波动程度。

下载: 导出CSV

参考文献(23)

[1]	Xingjian Shi, Zhourong Chen, Hao Wang, et al. Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting[C]// Proceedings of Advances in Neural Information Processing Systems 28. NeurIPS, 2015. doi: 10.1016/j.image.2019.115682
[2]	Pradhan R, Aygun R S, Maskey M, et al. Tropical cyclone intensity estimation using a deep convolutional neural network[J]. IEEE Transactions on Image Processing, 2018, 27(2): 692−702. doi: 10.1109/TIP.2017.2766358
[3]	Musinguzi A, Akbar M K, Fleming J G, et al. Understanding Hurricane Storm Surge Generation and Propagation Using a Forecasting Model, Forecast Advisories and Best Track in a Wind Model, and Observed Data—Case Study Hurricane Rita[J]. Journal of Marine Science and Engineering, 2019, 7(3): 77. doi: 10.1109/TMM.2018.2867742
[4]	Changming Dong, Guangjun Xu, Guoqing Han, et al. Recent Developments in Artificial Intelligence in Oceanography[J]. Ocean-Land-Atmosphere Research, 2022, 2022. doi: 10.1109/TIP.2018.2867740
[5]	Deo M C, Sridhar Naidu C. Real time wave forecasting using neural networks[J]. Ocean Engineering, 1998, 26(3): 191−203. doi: 10.1016/S0029-8018(97)10025-7
[6]	Mandal S, Prabaharan N. Ocean wave forecasting using recurrent neural networks[J]. Ocean Engineering, 2006, 33(10): 1401−1410. doi: 10.1016/j.oceaneng.2005.08.007
[7]	Mahjoobi J, Adeli Mosabbeb E. Prediction of significant wave height using regressive support vector machines[J]. Ocean Engineering, 2009, 36(5): 339−347. doi: 10.1016/j.oceaneng.2009.01.001
[8]	Etemad-Shahidi A, Mahjoobi J. Comparison between M5’ model tree and neural networks for prediction of significant wave height in Lake Superior[J]. Ocean Engineering, 2009, 36(15/16): 1175−1181.
[9]	Dixit P, Londhe S, Dandawate Y. Removing prediction lag in wave height forecasting using Neuro-Wavelet modeling technique[J]. Ocean Engineering, 2015, 93: 74−83. doi: 10.1016/j.oceaneng.2014.10.009
[10]	Alexandre E, Cuadra L, Nieto-Borge J C, et al. A hybrid genetic algorithm—extreme learning machine approach for accurate significant wave height reconstruction[J]. Ocean Modelling, 2015, 92: 115−123. doi: 10.1016/j.ocemod.2015.06.010
[11]	Kumar N K, Savitha R, Mamun A A. Regional ocean wave height prediction using sequential learning neural networks[J]. Ocean Engineering, 2017, 129: 605−612. doi: 10.1016/j.oceaneng.2016.10.033
[12]	Wang Wenxu, Tang Ruichun, Li Cheng, et al. A BP neural network model optimized by Mind Evolutionary Algorithm for predicting the ocean wave heights[J]. Ocean Engineering, 2018, 162: 98−107. doi: 10.1016/j.oceaneng.2018.04.039
[13]	Akbarifard S, Radmanesh F. Predicting sea wave height using Symbiotic Organisms Search (SOS) algorithm[J]. Ocean Engineering, 2018, 167: 348−356. doi: 10.1016/j.oceaneng.2018.04.092
[14]	Londhe S N, Shah S, Dixit P R, et al. A coupled numerical and artificial neural network model for improving location specific wave forecast[J]. Applied Ocean Research, 2016, 59: 483−491. doi: 10.1016/j.apor.2016.07.004
[15]	James S C, Zhang Yushan, O’Donncha F. A machine learning framework to forecast wave conditions[J]. Coastal Engineering, 2018, 137: 1−10. doi: 10.1016/j.coastaleng.2018.03.004
[16]	Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural Computation, 1997, 9(8): 1735−1780. doi: 10.1162/neco.1997.9.8.1735
[17]	Fan Shuntao, Xiao Nianhao, Dong Sheng. A novel model to predict significant wave height based on long short-term memory network[J]. Ocean Engineering, 2020, 205: 107298. doi: 10.1016/j.oceaneng.2020.107298
[18]	Zhou Shuyi, Xie Wenhong, Lu Yuxiang, et al. ConvLSTM-based wave forecasts in the South and East China Seas[J]. Frontiers in Marine Science, 2021, 8: 680079. doi: 10.3389/fmars.2021.680079
[19]	Gao Zhiyi, Liu Xing, Yv Fujiang, et al. Learning wave fields evolution in North West Pacific with deep neural networks[J]. Applied Ocean Research, 2023, 130: 103393. doi: 10.1016/j.apor.2022.103393
[20]	秦知朋, 陈永平, 潘毅, 等. 基于BO-LSTM神经网络模型的台风浪波高预报方法研究[J]. 海洋学报, 2024, 46(10): 107−116. Qin Zhipeng, Chen Yongping, Pan Yi, et al. Research on typhoon wave height prediction method based on BO-LSTM neural network model[J]. Haiyang Xuebao, 2024, 46(10): 107−116.
[21]	Ris R C, Holthuijsen L H, Booij N. A third-generation wave model for coastal regions: 2. Verification [J]. Journal of Geophysical Research: Oceans, 1999, 104(C4): 7667−7681.
[22]	邱立国, 梁小力, 陈斌, 等. 西沙群岛宣德环礁波浪分布特征[J]. 热带海洋学报, 2025, 44(5): 50−64. Qiu Liguo, Liang Xiaoli, Chen Bin, et al. Wave distribution characteristics of Xuande Atoll, Xisha Islands[J]. Journal of Tropical Oceanography, 2025, 44(5): 50−64.
[23]	Hersbach H, Bell B, Berrisford P, et al. ERA5 hourly data on single levels from 1940 to present[EB/OL]. Copernicus Climate Change Service (C3S) Climate Data Store (CDS) [2024−10−16]. https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=overview.