Fig. 1

Experimental flowchart"

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"

Fig. 3

Eddy detection and tracking algorithm process"

Tab. 1

Statistical table of RMSE of SLA prediction errors for multiple models"

Tab. 2

Statistical table for the number of eddy trackers"

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"

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"

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"

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"

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"