The frequency division prediction - optimal ensemble strategy improves the prediction ability of river dissolved oxygen

The accurate prediction of dissolved oxygen (DO) concentration in rivers is very important for the management of aquatic ecosystems, but the traditional model for predicting the nonlinear changes of dissolved oxygen in rivers is still insufficient. In this paper, a frequency division prediction framework based on the optimal ensemble of Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is proposed. The dissolved oxygen sequence was decomposed into multiple components by CEEMDAN, and the long short-term memory network (LSTM), support vector regression (SVR) and multi-layer perceptron (MLP) models were constructed to predict each component independently. The constrained grid search algorithm is introduced to realize the complementary advantages of each model through dynamic combination, and the optimal integration scheme is obtained with the goal of minimizing the mean absolute error (MAE) of the training set. The empirical study of monitoring sections A and B in the Ganjiang River Basin shows that : in the prediction task, the average absolute error of the integrated model is 28.1–30.3% lower than that of the optimal single model, the root mean square error ( RMSE ) is 27.4–35.3% lower, and the coefficient of determination ( R2 ) reaches 0.898 and 0.973. In particular, the error accumulation rate in the 3-day prediction is 27.7% lower than that of the traditional hybrid model. This framework enables the modes of multi-component dissolved oxygen series prediction to be effectively aliasing, and provides an extensible technical path for the intelligent management of the basin.