Abstract: Abstract: Dissolved oxygen (DO) has become an important parameter to predict water quality in modern aquaculture. The excessive or insufficient DO content in the water has posed a great impact on the metabolism and physiological functions of breeding organisms, even a threat to normal growth. However, DO is also characterized by nonlinear, large inertia, strong coupling, and time-varying, while easily affected by many factors, such as weather, water quality, and human activities. Therefore the DO prediction is highly demanding for the reduction and prevention of aquaculture disaster in safe agricultural production. In this study, a hybrid DO prediction model was proposed in aquaculture using K-means clustering and improved long short-term memory neural network (KLSTM). Meanwhile, the improved particle swarm optimization (IPSO) was introduced to optimize the parameter selection of the model, thereby predicting the change of DO in aquaculture from the perspective of time series. The similarity among variables was calculated, according to the weight of influencing factors under different weather conditions. K-means clustering was then applied to divide the dataset into multiple clusters. The periodic changes and trends of variables were determined to optimize the selection of training samples, while reducing the running time of the whole model. An LSTM-DO prediction model was then established for aquaculture using deep learning frameworks. The KLSTM-PSO prediction curve was closer to the actual one than others, indicating a higher prediction accuracy. The experimental results show that the prediction error of the model fluctuated less, while the prediction accuracy was higher in good weather conditions. The MAPE, RMSE, MAE, and NSC of the proposed model were 0.129 5, 0.645 3, 0.461 3, and 0.902 2, respectively, when the weather changed suddenly. The best NSC performance was found in the IPSO-KLSTM neural network among the six models. On August 8, the NSC of IPSO-KLSTM was 0.902 2, that of PSO-KLSTM was 0.887 6, that of PSO-LSSVM was 0.864 6, and that of LSTM, ELM and BP were 0.857 7, 0.826 6 and 0.829 8 respectively, indicating a higher prediction performance of the model with the optimized selection of samples and parameters. Similarly, the average RMSE of the model was improved by 17.10%, 24.89%, and 24.21%, respectively, compared with the single LSTM, ELM, and BP models, indicating a higher tolerance to the anomaly or missing sensor data caused by uncertain mixed weather conditions. Correspondingly, the LSTM neural network was an effective way to predict DO content in the intensive aquaculture, further to balance the stability and accuracy of prediction. Therefore, the IPSO-KLSTM was widely expected to serve as the optimal sample selection, with the low interference between different samples under the weather conditions, and high tolerance to emergencies for a higher prediction accuracy. The finding can offer a highly accurate prediction framework to track the dissolved oxygen in modern aquaculture.