Abstract: Water quality regulation is one of the most important tasks in intensive aquaculture management. Grasping the trend of the dissolved oxygen concentration timely and accurately and regulating water quality dynamics are the key for healthy growth in the non-stress environment of aquatic products in order to solve the low prediction accuracy, inferior capability of dynamic learning, online updates, and high computational complexity of the traditional online forecasting methods for water quality in intensive aquaculture. The online prediction model of dissolved oxygen content in intensive aquaculture eriocheir sinensis cultures was introduced, which was based on the least squares support vector machine (LSSVR) with time series similar data. The time series data collected online was segmented clustered using a feature points segmented time warping distance algorithm. The subsequence data sets reduced the size and optimized the LSSVR models training process, achieving multiple LSSVR models online modeling, and segmented memory and storage. According to the forecast data sequence and LSSVR sub-model similarity, it adaptively chose the optimal sub-model to get the predicted output. The online model was used for the prediction of the dissolved oxygen changing in high-density eriocheir sinensis culture ponds during July 21, 2012 to July 31, 2012 in Yixing City, Jiangsu Province, China. Experimental results showed that the proposed prediction model of FPSTWD-LSSVR had a better prediction effect than the FPSTWD-LSSVR, ILSSVR, SONB-LSSVR, or off-line LSSVR algorithms. Under the same experimental conditions, the relative mean absolute percentage error (MAPE), maximum relative error (Emax), relative root mean square error (RRMSE), and the running time differences between the FPSTWD-LSSVR and ILSSVR models were 47.93%, 43.47%%, 30.91%, and 5.16 s in the test period respectively. The relative MAPE, Emax, RRMSE, and the running time differences between the FPSTWD-LSSVR and SONB-LSSVR models were 39.99%, 33.43%, 22.40%, and 2.74 s in the test period respectively. It is obvious that FPSTWD-LSSVR is more accurate than ILSSVR and SONB-LSSVR. The relative MAPE, Emax, RRMSE and the running time differences between the FPSTWD-LSSVR and off-line LSSVR models were 16.14%, 9.03%, 8.41%, and 11.36 s in the test period respectively. The lower sample number, which cannot cover all types of characteristic in time series data, probably caused the prediction performance of FPSTWD-LSSVR to be slightly lower than the off-line LSSVR model. Overall, the online prediction model has a low computational complexity, fast convergence rate, high online prediction accuracy, and strong generalization ability. It is an effective online prediction method for the dissolved oxygen controlling in the high density eriocheir sinensis culture, and provides the basis of decisions for controlling water quality, setting the aquaculture water plan, and reducing the risk of cultivation.