基于PredRNN++模型对南海中尺度涡旋的预测研究
赵杰(1996—), 男, 安徽省铜陵市人, 硕士研究生, 主要从事人工智能海洋学研究。email: |
Copy editor: 孙翠慈
收稿日期: 2023-05-12
修回日期: 2023-06-08
网络出版日期: 2023-06-20
Prediction of mesoscale eddies in the South China Sea based on the PredRNN++ model
Copy editor: SUN Cuici
Received date: 2023-05-12
Revised date: 2023-06-08
Online published: 2023-06-20
基于26年的海表面高度异常、海表面风速异常、海表面温度异常资料, 利用时空序列预测模型PredRNN++, 本文预报 1~28d时效的南海中尺度涡旋轨迹和南海西部偶极子活动。结果表明, PredRNN++模型能从整体上考虑整个南海区域时空演变特征和环境风场、温度场的作用, 在短期(1~2周)、中期(3~4周)预报上具有良好的性能。该模型具备一定预报涡旋产生、消亡的能力, 且能将涡旋轨迹4周预报误差控制在42.1km, 对于生命时长小于100d的涡旋生命中期的位置、振幅预报误差小。此外模型在8—11月份的月平均、4天平均下的任意时间点和任意预报时效下均能较好地追踪到偶极子结构的演变、强度变化, 偶极子涡旋相关属性预报误差最小且存在年际、类型差异, 2017年涡旋1~4周振幅位置、预报、半径误差最小, 分别为40~60km、3~5cm、20~40km, 且气旋涡位置预报效果优于反气旋涡。
赵杰 , 林延奖 , 刘燃 , 杜榕 . 基于PredRNN++模型对南海中尺度涡旋的预测研究[J]. 热带海洋学报, 2024 , 43(1) : 16 -27 . DOI: 10.11978/2023060
Based on 26 years of data on sea level anomalies, sea surface wind speed anomalies, and sea surface temperature anomalies, using the spatiotemporal series prediction model PredRNN++, this paper predicts the trajectory of mesoscale eddies in the South China Sea and dipole activity in the western South China Sea over a period of 1 to 28 days. The results indicate that the PredRNN++model can comprehensively consider the spatiotemporal evolution characteristics of the entire South China Sea region and the role of environmental wind and temperature fields, and has good performance in short-term (1~2weeks) and medium-term (3~4weeks) forecasting. This model has the ability to predict the generation and disappearance of eddies to a certain extent, and can control the 4-cycle prediction error of eddy trajectories to 42.1 km. For eddies with a lifespan of less than 100 days, the mid-term position and amplitude prediction error are small. In addition, the model can better track the evolution and intensity change of dipole structure at any time point under the monthly average, 4-day average and any forecast time effect in August-November. The prediction error of dipole eddy related attributes is the smallest and there are interannual and type differences. In 2017, the amplitude position, prediction and radius error of eddy 1-4 cycles are the smallest, which are 40~60 km, 3~5 cm and 20~40 km respectively, and the prediction effect of cyclone position is better than that of anticyclone.
图2 a. 时空记忆单元的级联操作CasualLSTM结构; b. PredRNN++总体结构H、 C、 M、Z分别代表隐藏层状态、时间记忆、空间记忆、用于传递梯度的隐藏层状态 Fig. 2 (a) for the CasualLSTM structure, (b) represents H, C, M, and Z in the overall structure diagram of PredRNN++, representing the hidden layer state, temporal memory, spatial memory, and hidden layer state used to transfer gradients |
表1 多个模型涡旋位置预报误差RMSE统计表Tab. 1 Statistical table of RMSE of SLA prediction errors for multiple models |
模型名称 | 1周预报误差 (经向/纬向预报误差)/km | 2周预报误差 (经向/纬向预报误差)/km | 3周预报误差 (经向/纬向预报误差)/km | 4周预报误差 (经向/纬向预报误差)/km |
---|---|---|---|---|
回归模型 (Li et al, 2019) | — (32.7/29.5) | — (55.1 /47.3) | — (72.5 /61.4) | — (89.2 /73.5) |
LSTM | 36.5 (23.4/28.0) | 51.8 (34.8/38.4) | 68.7 (47.3/49.8) | 84.8 (58.5/61.4) |
Gradient boosting | 31.2 (22.5/21.6) | 59.1 (41.6/41.7) | 83.6 (57.8/60.1) | 104.3 (72.2/74.9) |
随机森林 | 51.3 (40.2/31.6) | 68.1 (51.8/44.0) | 86.9 (64.4/58.1) | 105.1 (76.6/71.5) |
Extra tree | 47.2 (38.0/27.8) | 69.1 (53.3/43.6) | 89.9 (67.0/59.6) | 108.8 (79.8/73.4) |
Extra tree (Wang et al, 2020) | — (28.8/23.8) | — (36.9/30.6) | — (41.9/34.1) | — (47.2/37.2) |
PredRNN++ | 28.1 (19.7/20.0) | 36.4 (25.6/25.8) | 40.8 (28.8/28.7) | 42.1 (29.3/30.0) |
注: — 表示无数据 |
表2 涡旋追踪个数统计表Tab. 2 Statistical table for the number of eddy trackers |
预报 时效 | 观测涡旋轨迹 追踪到个数 | 涡旋轨迹第一天被追踪到的涡旋轨迹个数 | 涡旋轨迹最后一天被 追踪到的涡旋轨迹个数 |
---|---|---|---|
1 | 549 | 297 | 381 |
2 | 549 | 239 | 388 |
3 | 549 | 211 | 367 |
4 | 549 | 181 | 357 |
5 | 549 | 165 | 338 |
6 | 548 | 153 | 316 |
7 | 548 | 129 | 319 |
14 | 528 | 71 | 192 |
21 | 482 | 66 | 110 |
28 | 426 | 47 | 82 |
图4 不同类型涡旋预报误差与预报时效、涡旋演变阶段(r)的关系图a. 涡旋生命周期不同时期(横轴)的不同预报时效下(颜色)的振幅; b. 位置预报误差RMSE(纵轴). 每个子图从左往右分别为E1、E2、E3三类涡旋, 不同颜色代表不同预报时效, 从1周到4周, 横轴为r值, 表征涡旋所处的生命阶段 Fig. 4 The relationship diagrams (a) and (b) between the prediction errors of different types of eddies and their prediction time and evolution stages show the RMSE of amplitude and position prediction errors (vertical axis) of different prediction times (colors) at different stages of the eddy life cycle (horizontal axis). Each subgraph represents three types of eddies, E1, E2, and E3, from left to right. Different colors represent different prediction times, from 1 to 4 weeks, and the horizontal axis represents the r value, representing the life stage of the eddy |
图5 2017年及2018年7—11月份越南东海岸偶极子的观测分布图和预报分布图a. 2017; b. 2018。每张子图自左往右对应观测分布图、预报时效1~4周的预报分布图, 自上而下对应7—11月份 Fig. 5 Observation and forecast distribution maps of the dipole on the east coast of Vietnam from July to November 2017 (a) and 2018 (b). Each sub map corresponds to the observation distribution map and the forecast distribution map with a prediction time of 1~4 weeks from left to right, and corresponds to July to November from top to bottom |
图6 2017、2018年8—11月份越南东海岸偶极子的4天平均下的追踪效果图(a)、(b)分别代表着2017、2018年, (a)自上而下依次对应8月29日到10月24日时间范围内4天间隔的时间点, (b)自上而下依次对应8月25日到10月4日时间范围内4天间隔的时间点, (a)、(b)子图中第1列为越南东海岸SLA及偶极子结构观测结果, 第2~5列代 Fig. 6 The 4-day average tracking effect of the dipole on the east coast of Vietnam from August to November 2017 and 2018, represented by (a) and (b), respectively, (a) corresponding to the 4-day interval between August 29 and October 24 from top to bottom, (b) corresponding to the 4-day interval between August 25 and October 4 from top to bottom, the first column in the sub figure (a) and (b) shows the observation results of SLA and dipole structure on the east coast of Vietnam, and the second to fifth columns represent the prediction results of SLA and dipole structure on the east coast of Vietnam under the 1~4 week forecast time effect. The black and red outlines are the edges of cyclones and anticyclone, respectively |
图7 2017、2018年8—11月份越南东海岸偶极子强度观测及预报值随着时间变化的折线图不同颜色的折线分别代表偶极子强度的观测值、1~4周预报值随着时间点的变化, (a)、(b)分别代表着2017、2018年, 横轴代表时间, 纵轴代表偶极子指数大小 Fig. 7 A line chart showing the changes of dipole intensity observation and forecast values along the east coast of Vietnam from August to November 2017 to November 2018. The lines in different colors represent the changes of dipole intensity observation values and 1~4 week forecast values along time points, (a) and (b) represent 2017 and 2018 respectively, the horizontal axis represents time, and the vertical axis represents dipole index size |
图8 2017(a—c)与2018(d—f)年偶极子结构中气旋涡与反气旋涡误差随着预报时效的变化图红色、蓝色折线分别代表反气旋涡、气旋涡, 图中自左向右各行分别为位置误差、振幅误差、半径误差的MAE值 Fig. 8 The change of errors of Mesocyclone and anticyclone in the dipole structure with the prediction time effect in 2017 (a, b, c) and 2018 (d, e, f). The red and blue broken lines represent anticyclone and cyclone, respectively, and the lines from left to right in the figure are MAE of position errors, amplitude errors and radius errors. The abscissa of each sub figure is the prediction time effect |
[1] |
韩玉康, 周林, 吴炎成, 2016. 基于HYCOM的南海中尺度涡数值模拟[J]. 海洋通报, 35(3): 299-316.
|
[2] |
李佳讯, 张韧, 陈奕德, 等, 2011. 海洋中尺度涡建模及其在水声传播影响研究中的应用[J]. 海洋通报, 30(1): 37-46.
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
/
〈 | 〉 |