Time space hybrid network XLTNET: A high-precision current velocity prediction model targeting local sea areas in the South Pacific
-
摘要: 精确的长期海表流速预测对海洋科学研究至关重要，但现有深度学习模型在处理高维时空序列时，普遍存在误差累积和长期稳定性不足的问题。为解决此挑战，本研究提出了一种创新的时空融合网络XLTNET。该模型基于编码器−解码器架构，其核心在于高效融合了两个关键模块：一个采用K近邻（KNN）稀疏自注意力机制的改进Swin Transformer，用于精准捕捉多尺度的空间动力学特征；以及一个扩展长短期记忆网络（xLSTM），用于增强对时间序列长程依赖的建模能力。实验基于哥白尼海洋服务中心的再分析数据集，选取了包含海表流速（U、V分量）、温度、盐度和高度在内的5个海洋要素。结果表明，XLTNET在长期预测任务中展现出卓越的性能和稳定性。在第15天的预测中，XLTNET是唯一在两个流向上R值均能保持在0.7以上的模型；其U向R值相比ASTMEN、ConvLSTM和LSTM分别提升了7.3%、18.0%和20.7%；V向R值分别提升了8.7%、15.6%和17.4%。消融实验进一步证实了模型各组件及深度融合策略的必要性。本研究为高精度的长期海表流速预测提供了一个性能优异的解决方案。Abstract: Accurate long-term forecasting of ocean current fluid is crucial for marine science research, yet existing deep learning models generally suffer from error accumulation and insufficient long-term stability when processing high-dimensional spatiotemporal sequences. To address this challenge, this study proposes an innovative spatiotemporal fusion network, XLTNET. The model is based on an encoder-decoder architecture, with its core lying in the efficient fusion of two key modules: an improved Swin Transformer that adopts the K-Nearest Neighbors (KNN) sparse self-attention mechanism for precisely capturing multi-scale spatial dynamics, and an extended Long Short-Term Memory network (xLSTM) for enhancing the modeling of long-range temporal dependencies. Experiments were conducted based on the reanalysis dataset from the Copernicus Marine Service, utilizing five ocean elements including ocean current fluid (U and V components), temperature, salinity, and height. The results demonstrate that XLTNET exhibits superior performance and stability in long-term forecasting tasks. In the 15-day forecast, XLTNET was the only model to maintain an R-value above 0.7 in both flow directions. Its U-direction R-value improved by 7.3%, 18.0%, and 20.7% compared to ASTMEN, ConvLSTM, and LSTM, respectively, while its V-direction R-value showed improvements of 8.7%, 15.6%, and 17.4%. Furthermore, ablation studies confirmed the necessity of each model component and the deep fusion strategy. This research provides a high-performance solution for high-precision, long-term ocean current fluid forecasting.
-
Key words:
- South Pacific /
- ocean current fluid prediction /
- Swin Transformer /
- spatiotemporal fusion /
- deep learning /
- xLSTM
-
图 1 南太平洋研究区域的地理位置信息
Fig. 1 Geographical information of the South Pacific research region
图 2 2017年10月15日研究区域的海表流速分布
Fig. 2 Ocean current fluid velocity distribution in the study area on October 15, 2017
图 4 xLSTM结构
sLSTM和mLSTM作为两种核心变体（左侧），其关键创新（如红色标记的指数门控）被整合进一个xLSTM块（中间）。多个xLSTM块通过残差连接堆叠，最终构成完整的xLSTM。图4改编自参考文献[27]
Fig. 4 xLSTM structure
sLSTM and mLSTM, as two core variants (on the left), have their key innovations (such as the exponential gating marked in red) integrated into an xLSTM block (in the middle). Multiple xLSTM blocks are stacked through residual connections, ultimately forming a complete xLSTM. Figure 4 is adapted from reference [27]
图 7 南太平洋海表流速时空融合预测模型结构
Fig. 7 South Pacific Ocean current spatiotemporal fusion prediction model structure
图 10 超参数配置热力图
热力图展示了核心参数（S，K，depth）对模型性能的影响
Fig. 10 Superparameter configuration heatmap
The heatmap illustrates the impact of core parameters (S, K, depth) on model performance
图 16 2020年11月30日(第15天)真实流场与预测流场
Fig. 16 Real and predicted flow fields on November 30, 2020 (Day 15)
表 1 sLSTM与mLSTM解码器资源消耗对比
Tab. 1 Comparison of resource consumption between sLSTM and mLSTM decoders
sLSTM mLSTM 单轮训练时长/h 22 30 平均显存占用/GB 17.56 23.58 平均显存占用率*/% 73 98 注：*显存占用率基于NVIDIA GeForce RTX 3090（24GB）显卡计算。 
下载: 导出CSV
表 2 4种模型预测结果U向MAE对比（单位：m/s）
Tab. 2 Comparison of U MAE for four model prediction results (unit: m/s)
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.06953 0.06955 0.06167 0.04861 3 0.07966 0.07994 0.07680 0.05913 5 0.09026 0.09920 0.11654 0.09574 7 0.09791 0.10769 0.12926 0.13772 10 0.10728 0.11729 0.14613 0.15729 12 0.11237 0.12222 0.16283 0.17391 15 0.11860 0.12827 0.18171 0.20136 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
表 4 4种模型预测结果U向RMSE对比（单位：m/s）
Tab. 4 Comparison of U RMSE for four model prediction results (unit: m/s)
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.07300 0.07070 0.06310 0.05800 3 0.09850 0.09930 0.10210 0.07660 5 0.11800 0.12230 0.13180 0.11990 7 0.12520 0.13280 0.14110 0.15520 10 0.13710 0.14570 0.17280 0.18240 12 0.14330 0.16300 0.20190 0.20560 15 0.15890 0.17820 0.23850 0.24440 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
表 5 4种模型预测结果V向RMSE对比（单位：m/s）
Tab. 5 Comparison of V RMSE for four model prediction results (unit: m/s)
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.08360 0.08020 0.07010 0.06510 3 0.10530 0.10690 0.09620 0.10420 5 0.11920 0.12910 0.13520 0.12880 7 0.12990 0.14250 0.17210 0.15520 10 0.14150 0.15520 0.19570 0.20140 12 0.15810 0.17140 0.22180 0.22930 15 0.16660 0.18920 0.24100 0.25620 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
表 6 4种模型预测结果U向R值对比
Tab. 6 Comparison of U R for four model prediction results
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.95560 0.95425 0.96381 0.97163 3 0.90190 0.89837 0.89752 0.91698 5 0.86656 0.85239 0.84815 0.86063 7 0.83692 0.81328 0.81506 0.80341 10 0.79301 0.75948 0.73168 0.72134 12 0.76381 0.72644 0.67349 0.66782 15 0.72906 0.67955 0.61876 0.60398 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
表 7 4种模型预测结果V向R值对比
Tab. 7 Comparison of V R for four model prediction results
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.95460 0.95202 0.96402 0.96863 3 0.90140 0.89203 0.90816 0.90134 5 0.86066 0.84338 0.82589 0.84381 7 0.82354 0.79966 0.78195 0.78806 10 0.77137 0.75948 0.71562 0.70498 12 0.74960 0.71614 0.67349 0.66498 15 0.70844 0.65195 0.61287 0.60361 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
表 3 4种模型预测结果V向MAE对比（单位：m/s）
Tab. 3 Comparison of V MAE for four model prediction results (unit: m/s)
天数/d XLTNET ASTMEN ConvLSTM LSTM 1 0.07163 0.07155 0.05853 0.05196 3 0.08194 0.08214 0.07685 0.08022 5 0.09527 0.09920 0.12761 0.09871 7 0.10889 0.11278 0.14812 0.13110 10 0.12279 0.12577 0.15112 0.16010 12 0.12637 0.13948 0.17319 0.18212 15 0.13261 0.15127 0.19752 0.20319 注：加粗数据为该天最佳模型表现。
下载: 导出CSV
-
[1] Hu Dunxin, Wu Lixin, Cai Wenju, et al. Pacific western boundary currents and their roles in climate[J]. Nature, 2015, 522(7556): 299−308. doi: 10.1038/nature14504 [2] Ganachaud A, Cravatte S, Melet A, et al. The Southwest Pacific Ocean circulation and climate experiment (SPICE)[J]. Journal of Geophysical Research: Oceans, 2014, 119(11): 7660−7686. doi: 10.1002/2013JC009678 [3] Roemmich D, Gilson J, Davis R, et al. Decadal spinup of the South Pacific subtropical gyre[J]. Journal of Physical Oceanography, 2007, 37(2): 162−173. doi: 10.1175/JPO3004.1 [4] 刘超. 热带太平洋海温准十年变率及其对西北太平洋热带气旋的影响[D]. 南京: 南京信息工程大学, 2022.Liu Chao. Quasi-decadal variability of sea surface temperature in the tropical pacific and its impact on tropical cyclones in the Northwest Pacific[D]. Nanjing: Nanjing University of Information Science and Technology, 2022. [5] 石世玮. 热带太平洋海表盐度年代际变化特征及其对ENSO的调制作用[D]. 南京: 南京信息工程大学, 2021.Shi Shiwei. Interdecadal variability of sea surface salinity in the Tropical Pacific and its modulation on ENSO[D]. Nanjing: Nanjing University of Information Science and Technology, 2021. [6] 冯志萍. 洪堡洋流生态系统结构及中上层重要经济渔业资源对气候变化的响应研究[D]. 上海: 上海海洋大学, 2024.Feng Zhiping. Response of ecosystem structure and pelagic important economic fishery resources in the Humboldt current system to climate[D]. Shanghai: Shanghai Ocean University, 2024. [7] Cai Wenju, Wang Guojian, Dewitte B, et al. Increased variability of eastern Pacific El Niño under greenhouse warming[J]. Nature, 2018, 564(7735): 201−206. doi: 10.1038/s41586-018-0776-9 [8] Van Sebille E, Wilcox C, Lebreton L, et al. A global inventory of small floating plastic debris[J]. Environmental Research Letters, 2015, 10(12): 124006. doi: 10.1088/1748-9326/10/12/124006 [9] Wang J L, Zhuang Hanqi, Chérubin L M, et al. Medium-term forecasting of loop current Eddy Cameron and Eddy Darwin formation in the Gulf of Mexico with a divide-and-conquer machine learning approach[J]. Journal of Geophysical Research: Oceans, 2019, 124(8): 5586−5606. doi: 10.1029/2019JC015172 [10] 黄奇华, 吴雄斌, 岳显昌, 等. 基于BP神经网络的高频地波雷达海流空间插值[J]. 海洋学报, 2019, 41(5): 138−145.Huang Qihua, Wu Xiongbin, Yue Xianchang, et al. Spatial interpolation of current mapped by HF surface wave radar using BP neural network[J]. Haiyang Xuebao, 2019, 41(5): 138−145. [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.Dong Shichao. Current prediction research based on ARIMA-BP neural network abstract[J]. China Science and Technology Information, 2014(2): 86−88. [12] Bayindir C. Predicting the ocean currents using deep learning[J]. TWMS Journal of Applied and Engineering Mathematics, 2023, 13(1): 373−385. [13] Zulfa I I, Novitasari D C R, Setiawan F, et al. Prediction of sea surface current velocity and direction using LSTM[J]. Indonesian Journal of Electronics and Instrumentation Systems, 2021, 11(1): 93−102. doi: 10.22146/ijeis.63669 [14] Immas A, Do N, Alam M R. Real-time in situ prediction of ocean currents[J]. Ocean Engineering, 2021, 228: 108922. doi: 10.1016/j.oceaneng.2021.108922 [15] Li Xin, Wang Fusheng, Song Tao, et al. ASTMEN: an adaptive spatiotemporal and multi-element fusion network for ocean surface currents forecasting[J]. Frontiers in Marine Science, 2023, 10: 1281387. doi: 10.3389/fmars.2023.1281387 [16] 袁红春, 管琦, 张天蛟. 基于LSTM-BP组合模型的南太平洋海流流速预测[J]. 海洋湖沼通报, 2022, 44(4): 10−16, doi: 10.13984/j.cnki.cn37-1141.2022.04.002Yuan Hongchun, Guan Qi, Zhang Tianjiao. South Pacific current forecast based on LSTM-BP neural network[J]. Transactions of Oceanology and Limnology, 2022, 44(4): 10−16, doi: 10.13984/j.cnki.cn37-1141.2022.04.002 [17] Thongniran N, Vateekul P, Jitkajornwanich K, et al. Spatio-temporal deep learning for ocean current prediction based on HF radar data[C]//Proceedings of the 2019 16th International Joint Conference on Computer Science and Software Engineering (JCSSE). Chonburi: IEEE, 2019: 254−259. [18] Liu Jingjing, Yang Jinkun, Liu Kexiu, et al. Ocean current prediction using the weighted pure attention mechanism[J]. Journal of Marine Science and Engineering, 2022, 10(5): 592. doi: 10.3390/jmse10050592 [19] De Bézenac E, Pajot A, Gallinari P. Deep learning for physical processes: incorporating prior scientific knowledge[J]. Journal of Statistical Mechanics: Theory and Experiment, 2019, 2019(12): 124009. doi: 10.1088/1742-5468/ab3195 [20] Xiao Changjiang, Chen Nengcheng, Hu Chuli, et al. Short and mid-term sea surface temperature prediction using time-series satellite data and LSTM-AdaBoost combination approach[J]. Remote Sensing of Environment, 2019, 233: 111358. doi: 10.1016/j.rse.2019.111358 [21] Shih S Y, Sun Fankeng, Lee H Y. Temporal pattern attention for multivariate time series forecasting[J]. Machine Learning, 2019, 108(8): 1421−1441. [22] Zhang Zhiyuan, Yan Yan, Wang Zihe, et al. SDPNet: a novel deeplearning method for ocean surface current prediction[J]. Journal of Physics: Conference Series, 2023, 2486(1): 012066. doi: 10.1088/1742-6596/2486/1/012066 [23] Panboonyuen T. SEA-ViT: forecasting sea surface currents using a vision transformer and GRU-based spatio-temporal covariance model[C]//Proceedings of the 2025 17th International Conference on Knowledge and Smart Technology (KST). Bangkok: IEEE, 2025: 1−6. [24] Cunha W, Viegas F, França C, et al. A comparative survey of instance selection methods applied to non-neural and transformer-based text classification[J]. ACM Computing Surveys, 2023, 55(13s): 265. [25] 余为, 温健. 洪堡洋流系统内重要渔业资源对气候和海洋环境变化的响应研究进展[J]. 上海海洋大学学报, 2022, 31(3): 620−630.Yu Wei, Wen Jian. Review on the response of important fishery resources to the climatic and environmental variability in the Humboldt Current System[J]. Journal of Shanghai Ocean University, 2022, 31(3): 620−630. [26] Greff K, Srivastava R K, Koutník J, et al. LSTM: a search space odyssey[J]. IEEE Transactions on Neural Networks and Learning Systems, 2017, 28(10): 2222−2232. doi: 10.1109/TNNLS.2016.2582924 [27] Beck M, Pöppel K, Spanring M, et al. xLSTM: extended long short-term memory[C]//Proceedings of the 38th International Conference on Neural Information Processing Systems. Vancouver: Curran Associates Inc. , 2025: 3417. [28] Alharthi M, Mahmood A. xLSTMTime: long-term time series forecasting with xLSTM[J]. AI, 2024, 5(3): 1482−1495. doi: 10.3390/ai5030071 [29] Liu Ze, Lin Yutong, Cao Yue, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 9992−10002. [30] Stramma L, Peterson R G, Tomczak M. The South Pacific current[J]. Journal of Physical Oceanography, 1995, 25(1): 77−91. doi: 10.1175/1520-0485(1995)025<0077:TSPC>2.0.CO;2 -
计量
- 文章访问数: 15
- HTML全文浏览量: 8
- PDF下载量: 1
- 被引次数: 0
