Abstract: Dissolved oxygen (DO) is one of the most crucial indicators of the water quality in aquaculture. The insufficient DO levels can inhibit the fish's feeding, halt growth, and potentially cause mortality. While excessive DO concentrations can cause the gas bubble disease. Existing DO prediction has focused mainly on the short- to medium-term in a range of 30–60 min (6–12 steps). Such limited early warning periods can only support the basic emergency interventions. It is often required to effectively manage the risks under the sudden fluctuations of the water quality. Furthermore, the most current DO prediction can rely on the conventional deep learning models, such as Transformers and long short-term memory (LSTM) networks. However, the high computational complexity can be confined to balance the long-term dependencies with the short-term dynamic patterns in the time-series data. In this study, a long-term DO prediction was proposed using an enhanced Mamba model with the mixed convolution block (MCB) and the frequency-enhanced channel attention (FECA) mechanism. The prediction horizon was extended to 120 min (24 steps). The emergency response time was expanded by 2-4 times. Aquaculture personnel were provided with a more sufficient time to handle abnormal DO fluctuations. The MCB was also introduced for the Mamba model's capacity to capture the short-term temporal patterns using depthwise separable convolutions with varying kernel sizes. Both long-term dependencies and short-term variations were more balanced in the water quality data. Additionally, the FECA was also designed to extract the frequency-domain features from the water quality time-series data. The weights were adaptively adjusted to emphasize the critical frequency information, thereby improving the detection of significant periodic patterns and high-frequency disturbances. Both time and frequency domain features were integrated into the optimization. The MCB-Mamba-FECA (MMFA) model was achieved in the complex temporal dependencies, significantly improving the accuracy of the long-term DO prediction. A real-world aquaculture dataset was collected from a fish farm in Nansha, Guangzhou Province, China. Experimental results demonstrated that the MMFA model performed better in the single-step predictions, compared with the existing DO prediction models. Substantial advantages were also found in the long-term forecasting tasks. Specifically, the MMFA model was achieved in the 24-step prediction scenario, with a mean absolute percentage error (MAPE) of 0.5223%, a ere root mean square error (RMSE) of 0.0726 mg/L, and a mean absolute error (MAE) of 0.0436 mg/L. There was the reductions of 26.37%, 14.29%, and 26.48%, respectively, compared with the best baseline model. Furthermore, the generalization and effectiveness of the MCB were enhanced in the temporal feature extraction of deep learning models, particularly in the multi-scale temporal patterns. Meanwhile, the average MAPE of the LSTM, temporal convolutional network (TCN), Transformer, and Mamba models decreased by 21.58%, 37.74%, 37.48%, and 25.17%, respectively, after MCB integration. Finally, the cross-seasonal experiments demonstrated that the MMFA exhibited superior generalization to maintain the high predictive accuracy over the rest DO models on the water quality datasets from different seasons. These findings can also provide the technological reference for the intelligent water quality in large-scale aquaculture.