Abstract: In order to improve the prediction accuracy of dissolved oxygen content, a complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is proposed and the artificial bee colony (ABC) is used to improve the long and short time memory network (LSTM) model for combined prediction of dissolved oxygen in water quality. First, CEEMDAN is used to calculate the dissolved oxygen content sequence, which is divided into several different natural mode components and trend components. After that, we apply the R/S class to compute the Hurst exponent H of different natural mode components and trend components, and according to the value of H, natural mode components and trend components are re-constructed into microscale, mesoscale and macroscale components, respectively. Finally, for three kinds of components, the ABC-LSTM model is used to predict three species scale components, and the linear weighted reconstruction method is used to obtain the final estimates of dissolved oxygen measured values. The proposed model is applied to the data acquisition system of Henglan water quality monitoring station in Xijiang River. The results show that the model can effectively improve the prediction accuracy of dissolved oxygen in Xijiang River, and the prediction accuracy is as high as 1.6978%. Compared with LSTM, support vector machine (SVM), the ABC-SVM and the artificial bee colony algorithm optimized back propagation neutral network (ABC-BPNN), the prediction accuracy is improved by respectively 2.2867%, 2.7544%, 2.3756% and 2.4448%, indicating that the model in this paper has higher accuracy than the traditional models, with better prediction performance and generalization ability, as well as lower error. This provides the basis for scientific decision-making for the Xijiang water quality monitoring management and maintenance.

Key words: CEEMDAN principle, empirical mode decomposition, artificial bee colony algorithm, long short memory network, dissolved oxygen