基于Holt-DI-EnKF的海洋温盐实时预测系统

张思文; 史文涛; 景连友; 涂楠; 魏成鹏

doi:10.11993/j.issn.2096-3920.2025-0149

基于Holt-DI-EnKF的海洋温盐实时预测系统

doi: 10.11993/j.issn.2096-3920.2025-0149

西北工业大学海洋研究院, 江苏太仓, 215400

基金项目: 国家自然科学基金项目 (62471397, 62371393); 西北工业大学太仓智汇港学生创新基金项目(TCCX240107).

详细信息

作者简介:
张思文(1999-), 女, 硕士在读, 主要研究方向为作用距离估计

通讯作者:
景连友(1986-), 男, 副教授, 主要研究方向为水声通信.

中图分类号: TJ630; P714+.1
计量
- 文章访问数: 22
- HTML全文浏览量: 14
- PDF下载量: 11
- 被引次数: 0
出版历程
- 收稿日期: 2025-10-22
- 修回日期: 2025-12-02
- 录用日期: 2025-12-11
- 网络出版日期: 2026-03-16

Holt-DI-EnKF-based System for Real-time Prediction of Ocean Temperature and Salinity

Ocean Institute, Northwestern Polytechnical University, Taicang 215400, China

摘要

摘要: 针对传统海洋温盐预测方法动态更新能力弱、不确定性量化依赖分布假设、数据同化与预测模型割裂的问题, 聚焦单点月平均温盐时间序列的实时预测任务, 构建兼顾精度、动态适应性与工程实用性的轻量化预测框架。文中集成并改进了经典的时间序列预测、不确定性估计与数据同化方法, 提出融合Holt二次指数平滑、Bootstrap置信区间与动态膨胀集合卡尔曼滤波(DI-EnKF)的预测框架, 即通过Holt模型分解温盐序列并优化参数, 利用Bootstrap量化预测不确定性, 并结合DI-EnKF同化实时观测数据修正误差, 形成“预测-同化”闭环。该框架在全球Argo温盐数据测试中, 温度预测精度显著优于差分自回归移动平均-长短期记忆网络(ARIMA-LSTM)、ARIMA-反向传播神经网络(ARIMA-BP)等混合模型, 盐度预测精度接近最优对比模型。
- 海洋温盐预测 /
- Holt二次指数平滑 /
- Bootstrap置信区间 /
- 动态膨胀集合卡尔曼滤波 /
- 数据同化
Abstract: To address the limitations of traditional ocean temperature and salinity prediction methods—weak dynamic update capability, distribution-dependent uncertainty quantification, and disjointed data assimilation-prediction models—this study focuses on real-time prediction of monthly average temperature and salinity time series at a single point. Aiming to develop a lightweight framework balancing accuracy, dynamic adaptability, and engineering practicality, it integrates and advances classic time series prediction, uncertainty estimation, and data assimilation methods, proposing a hybrid framework of Holt's double exponential smoothing, Bootstrap confidence intervals, and Dynamically Inflated Ensemble Kalman Filter(DI-EnKF). Specifically, the Holt model decomposes sequences and optimizes parameters, Bootstrap quantifies prediction uncertainty, and DI-EnKF assimilates real-time Argo observations to correct errors, forming a"prediction-assimilation"closed loop. Tests on global Argo data demonstrate that the framework outperforms hybrid models such as ARIMA-LSTM and ARIMA-BP in temperature prediction, with salinity prediction accuracy close to the optimal comparison model.
- Ocean temperature and salinity prediction /
- Holt's double exponential smoothing /
- Bootstrap confidence interval /
- Dynamically Inflated Ensemble Kalman Filter /
- Data assimilation

HTML全文

图 1 Argo浮标全球分布图

Figure 1. Global distribution map of Argo buoys

下载: 全尺寸图片幻灯片

图 2 时间选择界面

Figure 2. Time selection interface

下载: 全尺寸图片幻灯片

图 3 温度剖面图

Figure 3. Temperature profile diagram

下载: 全尺寸图片幻灯片

图 4 盐度剖面图

Figure 4. Salt profile diagram

下载: 全尺寸图片幻灯片

图 5 DI-EnKF同化流程图

Figure 5. Flow chart of DIEnKF assimilation

下载: 全尺寸图片幻灯片

图 6 温度预测结果

Figure 6. Prediction results of temperature

下载: 全尺寸图片幻灯片

图 7 盐度预测结果

Figure 7. Prediction results of salt

下载: 全尺寸图片幻灯片

表 1 温度预测误差对比

Table 1. Comparison of temperature prediction errors

模型	MAE/℃	RMSE/℃	MAPE/%
Holt-DI-EnKF	0.229	0.318	1.05
ARIMA(1,1,0)+LSTM	2.154	2.867	9.87
ARIMA(0,1,1)+LSTM	2.478	3.245	11.34
ARIMA(1,1,1)+LSTM	2.301	3.012	10.54
ARIMA(1,1,0)+BP	1.987	2.645	9.12
ARIMA(0,1,1)+BP	2.112	2.812	9.68
ARIMA(1,1,1)+BP	2.045	2.732	9.38

下载: 导出CSV

表 2 盐度预测误差对比

Table 2. Comparison of salt prediction errors

模型	MAE/‰	RMSE/‰	MAPE/%
Holt-DI-EnKF	0.088	0.094	0.25
ARIMA(1,1,0)+LSTM	0.091	0.101	0.26
ARIMA(0,1,1)+LSTM	0.095	0.105	0.27
ARIMA(1,1,1)+LSTM	0.089	0.099	0.25
ARIMA(1,1,0)+BP	0.103	0.112	0.29
ARIMA(0,1,1)+BP	0.107	0.116	0.30

下载: 导出CSV

表 3 温度预测消融实验误差对比

Table 3. Comparison of experimental errors in temperature prediction ablation experiments

区域	模型	MAE (℃)	RMSE (℃)	MAPE (%)
52°N,163°E	Holt	0.2413	0.5577	7.61
	Holt+EnKF	0.0349	0.0462	1.02
	Holt-DI-EnKF	0.0338	0.0447	0.99
42°S,56°E	Holt	1.2946	1.3327	10.99
	Holt+EnKF	0.0870	0.1173	0.80
	Holt-DI-EnKF	0.0357	0.0443	0.33
52°N,163°W	Holt	0.1738	0.4919	2.87
	Holt+EnKF	0.0184	0.0252	0.43
	Holt-DI-EnKF	0.0184	0.0247	0.43
58°S, 160°W	Holt	0.6102	0.6516	25.83
	Holt+EnKF	0.1150	0.1234	4.91
	Holt-DI-EnKF	0.1118	0.1201	4.79

下载: 导出CSV

表 4 盐度预测消融实验误差对比

Table 4. Comparison of experimental errors in salt prediction ablation experiments

区域	模型	MAE(‰)	RMSE(‰)	MAPE (%)
52°N,163°E	Holt	0.0331	0.0460	0.10
	Holt+EnKF	0.0305	0.0408	0.09
	Holt-DI-EnKF	0.0303	0.0460	0.09
42°S,56°E	Holt	0.6780	0.7285	1.94
	Holt+EnKF	0.0338	0.0436	0.11
	Holt-DI-EnKF	0.0383	0.0431	0.11
52°N,163°W	Holt	0.0399	0.0742	0.12
	Holt+EnKF	0.0319	0.0555	0.12
	Holt-DI-EnKF	0.0318	0.0552	0.10
58°S,160°W	Holt	0.0248	0.0265	0.07
	Holt+EnKF	0.0241	0.0261	0.07
	Holt-DI-EnKF	0.0243	0.0262	0.07

下载: 导出CSV

表 5 单点温盐预测各模块计算耗时

Table 5. Computation time of each module for single-point temperature/salinity prediction

	模块	所需时间/s
温度	Holt	19.12
	Holt-DI-EnKF	0.08
	其余模块	6.97
	总计	26.17
盐度	Holt	19.35
	Holt-DI-EnKF	0.09
	其余模块	7.1
	总计	26.54

下载: 导出CSV

参考文献(23)

[1]	王宁. 面向机器学习的海洋环境数据分析与预测研究[D]. 唐山: 华北理工大学, 2020: 3-4
[2]	张建华. 海温预报知识讲座第一讲海水温度预报概况[J]. 海洋预报, 2003, 20(4): 81-85.
[3]	Liu H L, Lin P F, Zheng W P, et al. A global eddy-resolving ocean forecast system in China-LICOM Forecast System(LFS)[J]. Journal of Operational Oceanography, 2023, 16(1): 15-27. doi: 10.1080/1755876X.2021.1902680
[4]	金向泽, 张学洪. 温盐环流与全球增暖的数值模拟 (一)纬向平均温盐环流的模拟[J]. 大气科学, 1994, 18(z1): 769-779. Jin X Ze, Zhang X H. Simulation of thermohaline circulation and global warming──(I)simulation of zonal mean thermohaline circulation[J]. Chinese Journal of Atmospheric Sciences, 1994, 18(z1): 769-779.
[5]	杜扬帆, 伍孝飞, 乔百友. 基于XGBoost-PredRNN++的海表面温度预测[J]. 计算机系统应用, 2022, 31, (10): 236-244. Du Y F;Wu X F;Qiao B Y. Sea surface temperature prediction based on XGBoost-PredRNN[J]. Computer Systems & Applications, 2022, 31, (10): 236-244.
[6]	桑燕芳, 王中根, 刘昌明. 水文时间序列分析方法研究进展[J]. 地理科学进展, 2013, 32(1): 20-30. SANG Y F, WANG Z G, LIU C M. Research progress on the time series analysis methods in hydrology[J]. Progress in Geography, 2013, 32(1): 20-30.
[7]	Evensen G. The Ensemble Kalman filter: Theoretical formulation and practical implementation[C]//Conference on Seminar on Recent Developments in Data Assimilation for Atmosphere and Ocean, 2004: 221-264.
[8]	沈浙奇, 唐佑民, 高艳秋. 集合资料同化方法的理论框架及其在海洋资料同化的研究展望[J]. 海洋学报, 2016, 38(3): 1-14. Shen Z Q, Tang Y M, Gao Y Q. The theoretical framework of the ensemble-based data assimilation method and its prospect in oceanic data assimilation[J]. Acta Oceanologica Sinica, 2016, 38(3): 1-14.
[9]	李亚蒙, 丁军航, 孙宝楠, 等. BP和RBF神经网络应用于海表温盐短期预测效果对比[J]. 海洋科学进展, 2022, 40(2): 220-232. doi: 10.12362/j.issn.1671-6647.2022.02.006 LI Y M, DING J H, SUN B N, et al. Comparison of short-term prediction effects of the sea surface temperature and salinity based on BP and RBF neural network[J]. Advances in Marine Science, 2022, 40(2): 220-232. doi: 10.12362/j.issn.1671-6647.2022.02.006
[10]	高国栋, 张文孝, 慕光宇. RBF网络和BP网络在海水盐度建模中的比较研究[J]. 海洋通报, 2011, 30(1): 12-15. doi: 10.3969/j.issn.1001-6392.2011.01.003 GAO G D, ZHANG W X, MU G Y. A comparative study on RBF network and BP network in the model of salinity[J]. Marine Science Bulletin, 2011, 30(1): 12-15. doi: 10.3969/j.issn.1001-6392.2011.01.003
[11]	董世超. 基于ARIMA-BP神经网络模型海流流速预测研究[J]. 中国科技信息, 2014(2): 86-88. Chao D S. Current Prediction Research Based on ARIMA-BP Neural Network Abstract[J]. China Science and Technology Information, 2014(2): 86-88.
[12]	胡泽煜. 基于非平稳时间序列的海面温度预测研究[D]. 上海海洋大学, 2021: 4, 31.
[13]	贺琪, 胡泽煜, 徐慧芳等. 基于经验模态分解-门控循环模型的海表温度预测方法[J]. 激光与光电子学进展, 2021, 58, (24): 334-342. He Q, Hu Z Y, Xu H F, et al. Sea surface temperature prediction method based on empirical mode decomposition-gated recurrent unit model[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2415005
[14]	Yang Y T, Dong J Y, Sun X, et al. A CFCC-LSTM model for sea surface temperature prediction[J]. IEEE Geoscience and Remote Sensing Letters, 2018, 15(2): 207-211. doi: 10.1109/LGRS.2017.2780843
[15]	Healy M J R. Smoothing, forecasting and prediction of discrete time series[J]. Journal of the Royal Statistical Society: Series A(Statistics in Society), 1964, 127(2): 292-293. doi: 10.2307/3150192
[16]	Holt C C. Forecasting seasonals and trends by exponentially weighted moving averages[J]. International Journal of Forecasting, 2004, 20(1): 5-10.
[17]	Winters P R. Forecasting sales by exponentially weighted moving averages[J]. Management Science, 1960, 6(3): 324-342. doi: 10.1287/mnsc.6.3.324
[18]	Hyndman R J, Koehler A B, Ord J K, et al. Prediction intervals for exponential smoothing using two new classes of state space models[J]. Journal of Forecasting, 2005, 24(1): 17-37. doi: 10.1002/for.938
[19]	Bergmeir C, Hyndman R J, Benitez J M. Bagging exponential smoothing methods using STL decomposition and Box-Cox transformation[J]. International Journal of Forecasting, 2016, 32(2): 303-312. doi: 10.1016/j.ijforecast.2015.07.002
[20]	张雅乐, 俞永强, 段晚锁. 四个耦合模式ENSO后报试验的“春季预报障碍”[J]. 气象学报, 2012, 70(3): 506-519. ZHANG Y L, YU Y Q, DUAN W S. The spring prediction barrier of ENSO in retrospective prediction experiments as shown by the four coupled ocean-atmosphere models[J]. Acta Meteorologica Sinica, 2012, 70(3): 506-519.
[21]	Pauthenet E, Bachelot L, Balem K, et al. Four-dimensional temperature, salinity and mixed-layer depth in the Gulf Stream, reconstructed from remote-sensing and in situ observations with neural networks[J]. Ocean Science, 2022, 18(4): 1221-1244. doi: 10.5194/os-18-1221-2022
[22]	Efron B. Bootstrap Methods: Another Look at the Jackknife[J]. The Annals of Statistics, 1979, 7(1): 1-26. doi: 10.1214/aos/1176344552
[23]	Samal U, Kumar A. Improving software reliability: a hybrid ARIMA-LSTM approach for fault prediction[J]. International Journal of System Assurance Engineering and Management, 2025: 1-13.