Data stream prediction model for dissolved oxygen in aquaculture water using PC-RELM

Dissolved oxygen (DO) is one of the most important parameters for the water quality in aquaculture water. Long-term low oxygen environment can dominate the growth and reproduction of fish. Hypoxia can also cause large areas of fish death. Accurate and efficient DO prediction and control strategies can improve aquaculture production efficiency for the fewer aquaculture risks. However, an effective DO prediction has always been a tough challenge in aquaculture, due to the interference of external weather and the DO complexity. Multi-source or single sensors are generally used to build the prediction models, without considering the DO characteristics under similar weather conditions. Particularly, there is an outstanding diurnal variation in the DO content. Moreover, some redundant data can be collected from the water quality sensors in automatic weather stations. In this study, the principal component analysis and clustering method optimized regularized extreme learning machine (PC-RELM) was proposed to realize the DO prediction, considering the influence of external weather conditions on the DO and the diurnal variation. Firstly, the principal component analysis (PCA) was applied to determine the most influencing factors on the DO concentration, and reduce the data dimension of the prediction model for the high efficiency of prediction; Secondly, the entropy weight method was utilized to calculate the weather environment index at different time points. Fast dynamic time warping (FastDTW) was used to measure the similarity of weather environment in the time series data streams; Then, the K-means algorithm was used to cluster the similarity of the time series using the weather environment index. And the sub-prediction models of regularized extreme learning machine (RELM) were constructed using the clustered datasets to forecast the DO concentration. Finally, the PC-RELM model was applied to the intelligent control process of DO in the aquaculture pond of the Wuxi Nanquan experimental base. The test results showed that the root-mean square error (RMSE) of PC-RELM prediction was 0.961 9, which outperformed the partial least squares optimized ELM (PLS-ELM), Least Square Support Vector Machine (LSSVM), and BP algorithms by 41.54%, 54.58%, and 67.16%, respectively. The mean square error (MSE) value of PC-RELM was 0.6941, which outperformed the PLS-ELM, LSSVM and BP algorithms by 46.26%, 59.98%, and 69.90%, respectively. Meanwhile, the Nash-Sutcliffe efficiency coefficient of PC-RELM was 0.712 8, which was much higher than the rest prediction. In addition, the PC-RELM presented a high running speed of 0.316 2 s. The efficiency of PC-RELM was improved by about 7, 10, and 40 times, respectively, compared with the PLS-ELM, LSSVM, and BP. The improved model can be expected to extract the change patterns of DO under different weather conditions, indicating high prediction accuracy and efficiency. The finding can provide high-quality data and theoretical support for the precise control of DO in the pond water.