Abstract: Argo (Array for Real-time Geostrophic Oceanography) is one of the main sources of marine environmental information, which can obtain a large range of global upper ocean surface temperature profile data, through automatic profiling buoy, satellite positioning, data assimilation and other technologies. In this paper, ConvGRU (Convolutional Gate Recurring Unit) was used as the prediction model of Argo temperature, and part of the Northwest Pacific Ocean was taken as the study area, and Argo data from 2004 to 2018 were selected as the training data. The horizontal sections at the depth of 0 m, 50 m, 100 m, 200 m and 300 m in 2019 are predicted and analyzed. The results show that the ConvGRU model has a good ability to simulate the variation trend of Argo temperature data. The RMSE (Root Mean Squared Error) of the training set and the verification set of the prediction model were 0.0462 ℃ and 0.0463 ℃ respectively, and the MAE(Mean Absolute Error) were 0.0442 ℃ and 0.0450 ℃ respectively. Its Acc (Accuracy) is above 99%; For the prediction assessment, the error range of RE (Relative Error) is between 0.0228-0.0427, indicating that the spatial characteristics of the predicted variation are in good agreement with the real value.