Abstract: Dissolved oxygen (DO) concentration is one of the critical indicators to assess the aquaculture water. It is essential to maintain the normal physiological activities of aquatic organisms, such as fish and shrimp. The variation in the DO concentration can directly impact the growth, development, metabolic activities, and survival rates of the cultured organisms. Excessively high concentrations can cause the gas bubble disease in fish, due to the formation of gas bubbles in their tissues. Conversely, the low concentrations can impair the fish growth rates, leading to the fish mortality due to a floating head. Accurate and effective prediction of the DO can timely monitor the water quality trends, in order to reduce the aquaculture risks for the high efficiency of aquaculture. However, the dynamic variation in the DO concentration can be dominated by multiple complex factors in practical aquaculture environments. There are significant non-linear and multi-scale features. An accurate and effective prediction model is often required for the DO. In this study, a FRS-BiGRU DO prediction model was proposed to integrate the feature information, residual network (ResNet), and squeeze-and-excitation networks (SENet) improved bidirectional gated recurrent unit (BiGRU). Firstly, water environmental characteristic quantities were constructed according to the time-delay coupling relationship between water temperature and DO. The characteristic interaction was analyzed to enhance the representativeness of the original sample set. Secondly, ResNet was utilized to extract the features from the DO time series data. A systematic investigation was also made on the correlations between various factors and DO using residual connections and deep information transfer. The SENet channel attention mechanism was integrated into ResNet in order to further enhance the key features. As a result, the dynamic adaptive adjustment of the channel importance was achieved to highlight some features with significant impacts on the prediction performance. Finally, the strategy was introduced into the BiGRU prediction model. The FRS-BiGRU DO prediction model was then constructed to capture the forward and backward information of the DO time series. The improved model was also applied to two ponds in Nanquan Aquaculture Experimental Base in Wuxi City, Jiangsu Province, China. The performance of the DO prediction was verified after the test. The experimental results demonstrated that the FRS-BiGRU model achieved the mean values of 0.1098, 0.0806, and 0.9970 for root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) in the two ponds, respectively. The improved model reduced RMSE by 59.91% and 49.42%, respectively, compared with the current mainstream models, including long short-term memory network (LSTM) and gated recurrent unit (GRU). Similarly, there were reductions in the MAE by 59.38% and 51.79%, respectively, indicating the high accuracy of the performance. The model proposed in this study demonstrates good predictive accuracy and can provide a theoretical basis for the early warning and regulation of DO in pond aquaculture.