Dissolved oxygen (DO) in water is an important ecological factor for the healthy growth of aquaculture products. Dissolved oxygen in ponds is susceptible to many factors, which would cause differences in temporal and spatial distribution. Most of the existing dissolved oxygen prediction methods are time series predictions for a single monitoring point, and it cannot describe the spatial distribution of dissolved oxygen in the pond. Therefore, it is very important to predict the spatial and temporal dissolved oxygen in ponds. A spatio-temporal prediction method of dissolved oxygen in ponds based on autoregressive recurrent neural network (DeepAR) and regularized extreme learning machine (RELM) was proposed. Firstly, according to the sample entropy (SE) of the original dissolved oxygen sequence of each monitoring point and the maximum mutual information coefficient (MIC) between the sequences, a monitoring point with a smaller entropy value and a greater correlation with each point was selected as the central monitoring point, and the pond spatial coordinate system was established with the central monitoring point as the origin. Secondly, the DeepAR algorithm was used to predict the time series of dissolved oxygen in the central monitoring point. Finally, the RELM algorithm was used to construct the spatial mapping relation model between the central monitoring point and the dissolved oxygen in each location of the pond, and the spatial prediction of the dissolved oxygen in the pond in the future was realized by combining the predicted value of the time series of the dissolved oxygen at the central monitoring point and the spatial coordinates of the pond. This method not only improved the accuracy of time series prediction, but also realized the spatial prediction of dissolved oxygen in ponds. Tested on a real dataset predicting the dissolved oxygen value of the pond space in the next 24 hours, the root mean square error (RMSE) was 1.2633mg/L, the average absolute error (MAE) was 0.9755mg/L, and the average absolute percentage error (MAPE) was 14.8732%. Compared with common prediction methods, the performance of each evaluation index was greatly improved, which could more accurately realize the spatio-temporal prediction of dissolved oxygen in ponds.