Ocean thermocline prediction method based on deep space-time residual network

The invention discloses an ocean thermocline prediction method based on a deep space-time residual network. The method is based on special attributes of spatio-temporal data. According to the invention, the closeness, periodicity and tendency of ocean thermocline evolution are simulated by using a residual neural network framework, for each attribute, a branch of each residual convolution unit isdesigned, modeling is carried out on the space attributes of the ocean temperature in each unit, an ST-ResNet dynamically aggregates the outputs of three residual neural networks according to the data, and different weights are distributed to different branches so as to predict the final thermocline condition of a certain specified sea area. Experiments of sea areas with the longitude of 95-degreeW to 115-degree W and the latitude of 9.5-degree N to 9.5-degree S show that compared with a traditional SVM method, the ST-ResNet provided by the invention has the advantages that under the same operation environment and dif